04 / Publications · Bibliographic evidence

Publication Radar.

Track AI research across preprints, conferences, and journals.

New papers. Publication updates. One connected research record.Subscribe via RSS ↗

Latest weekly report

Publication Radar Weekly

New research and publication milestones, with source coverage in view.

Read the report

Follow the people behind the papers.Explore AI scientists by field →

Reset
Topics, venue and dates

All papers sort by earliest public release. Record-type and venue views use the qualifying record’s publication date, or an explicit acceptance date. Partial dates retain their precision; undated records appear last.

Research feed

2,569 distinct works · Page 1 · Release b4cfb7c2

preprintarxiv

CAFIN: CUDA-accelerated Adaptive Fisher Information Matrix Network Optimization via GPU-Native Kernel Fusion

Maryam Arshad, Muhammad Khuram Shahzad

Maryam Arshad, Muhammad Khuram Shahzad

First announced by arXiv: 2026-11-13

learning optimizationreasoning agents
Abstract from arxiv

Second-order optimizers improve deep neural network training using curvature information, but their practical efficiency is limited by GPU execution overhead. AdaFisher reduces Fisher Information Matrix (FIM) approximation costs through diagonal Kronecker structures and exponential moving averages, yet its PyTorch implementation requires sequential CUDA kernel launches, intermediate tensor allocations, and repeated memory accesses. We introduce CAFIN, a GPU-native CUDA kernel fusion framework that improves AdaFisher execution efficiency without changing its optimization formulation. The proposed approach progressively integrates FIM preconditioner construction, exponential moving average updates, gradient preconditioning, momentum accumulation, weight decay, and parameter updates into fused CUDA kernels. Shared-memory reuse, register-resident computation, and in-place operations further reduce memory traffic and intermediate allocations. Experiments using ResNet-18 on CIFAR-10, CIFAR-100, and Tiny ImageNet across NVIDIA GTX 1660 Ti and RTX 3060 Laptop GPUs demonstrate up to 78.4% lower isolated FIM construction latency and 2.1-26.9% lower optimizer-step latency. The implementation also eliminates approximately 18.91 MB of intermediate GPU memory allocations per optimizer step. Numerical validation confirms consistency within the predefined FP32 tolerance, while convergence accuracy remains comparable across evaluated datasets and random seeds. These findings demonstrate that CUDA kernel fusion can substantially accelerate second-order optimization while preserving AdaFisher's underlying mathematical formulation and learning behavior.

preprintarxiv

An Asynchronous Federated Learning Platform for Lung Adenocarcinoma Grading via Foundation Models and Structural Pruning

Moon-Il Joo, Md Ariful Islam Mozumder, Hee-Cheol Kim

Moon-Il Joo, Md Ariful Islam Mozumder, Hee-Cheol Kim

First announced by arXiv: 2026-11-13

foundation modelslearning optimization
Abstract from arxiv

Classification of lung adenocarcinoma into poorly differentiated (PD) and well-to-moderately differentiated (WMD) subtypes plays an important role in patient survival; however, consistent classification is difficult due to inter-observer variability among pathologists. Collecting data from multiple hospitals is also challenging due to patient privacy concerns. This study proposes a web-based federated learning system that shares only locally trained model weights without sharing patient data between institutions. The UNI2 foundation model and CLAM-SB were used to classify pathological whole slide images. A four-client federated learning system was constructed, and the NLST cohort was reserved exclusively for external validation. In centralized evaluation, joint training with CPTAC and TCGA outperformed single-institution training on external validation, confirming the necessity of multi-institutional data integration. FedAvg, FedProx, FedBN, and FedNova were compared as federated learning algorithms under identical experimental conditions. Test AUC ranked as FedBN, FedProx, FedNova, and FedAvg. FedNova showed the most balanced classification performance between PD and WMD. L1-norm structural pruning at 20%, 30%, and 40% sparsity was applied to reduce model transmission size. FedAvg, FedProx, and FedBN showed a tendency of decreasing Test AUC under pruning, while FedNova achieved improved Test AUC across all sparsity levels. A web-based application was developed supporting slide upload, local training, AI classification, and attention heatmap-based visualization. The combination of FedNova and 40% pruning was confirmed as a practical approach that improves communication efficiency and classification performance while protecting patient privacy.

preprintarxiv

Adversarial Risk in the Generative AI Era Necessitates Dropping the Small Epsilon Ball

Andrew C. Cullen, Neil G. Marchant, Paul Montague +2 authors

Andrew C. Cullen, Neil G. Marchant, Paul Montague, Jiani Xie, Benjamin I. P. Rubinstein

First announced by arXiv: 2026-11-13

learning optimization
Abstract from arxiv

Adversarial Machine Learning research is currently centered on the search for the epsilon-ball: minimal perturbations designed to fool models while remaining imperceptible to humans. We argue that this fixation on $\ell_p$ norms is a metrological trap that communicates a distorted perception of risk to defenders, one that is unaligned with real-world system vulnerabilities. This paradigm misleads defenders into optimizing for narrow, mathematical robustness at the expense of systemic security, while simultaneously incentivizing attackers to exploit non-overlapping alternative adversarial pathways. Ultimately, we argue that the dominance of the $\ell_p$ ball does not just fail to reduce harm---it may actively facilitate it by communicating a false sense of security against the wrong threats.

preprintarxiv

StructRL: Model-Free Reinforcement Learning with Replay Graphs

Ivo Nowak

Ivo Nowak

First announced by arXiv: 2026-11-13

learning optimization
Abstract from arxiv

StructRL turns the experience collected during reinforcement learning into an explicit replay graph. Once the graph contains a path to the task goal, dynamic programming selects actions on the observed part of the state space; these actions form a graph policy that is defined only where the required transitions have been observed. StructRL transfers these actions to model-free learners: as action preferences for a DQN Q-model and as target directions along shortest observed paths for a SAC actor. For DQN on three sparse-reward MiniGrid tasks, final success rises from 0.10 to 0.90 (DoorKey), 0.10 to 0.80 (KeyCorridor) and 0 to 0.50 (ObstructedMaze); plain DQN does not catch up with three times the budget. For SAC on continuous PointMaze, graph-derived direction labels raise final success from 0.17 to 1.0 on seven matched runs that reached the goal. Distilling the graph policy into a fresh Q-model gives a DoorKey 6x6 controller that reaches the goal from 91% of reachable states, against 38% for plain DQN. Scaling experiments identify three bottlenecks: incomplete graph support, structural aliasing, and loss of graph information in high-dimensional inputs.

preprintarxiv

KernelMosaic: An Open Dataset and Agentic Framework for CUDA Kernel Generation

Ningyuan Xi, Yuhang Lai, Ning Miao

Ningyuan Xi, Yuhang Lai, Ning Miao

First announced by arXiv: 2026-11-13

learning optimizationreasoning agents
Abstract from arxiv

Training large language models to generate efficient CUDA kernels requires supervision spanning diverse computational requirements and iterative code optimization. Constructing this supervision is challenging because of heterogeneous specifications, and useful implementations emerge through iterative complicated testing and refinement. We introduce KernelMosaic, an open dataset pairing diverse CUDA tasks with verified solutions and trajectories. Human-guided review and diversity-aware selection curate tasks expressed as natural-language descriptions, reference programs, mathematical specifications, and their intersections. To build the solutions and trajectories, our agentic framework guides a strong LLM through implementation, correctness repair, and performance optimization with reusable CUDA guidance and execution feedback. Supervised fine-tuning of a small open-source model on KernelMosaic raises correctness from 4%, 0%, and 0% to 88%, 82%, and 16% on KernelBench L1, L2, and L3, respectively, while increasing the fraction of tasks with kernels faster than PyTorch eager at every level. Further training experiments show that supervision from the curated trajectories substantially outperforms direct responses and promotes sustained testing and revision. These results demonstrate that KernelMosaic enables a general-purpose model to achieve substantial gains in CUDA kernel generation.

preprintarxiv

FreeDeshadow: Shadow Removal for User-Specified Objects With Multimodal Prompts

Zhiliang Wu, Jihao Gu, Kun Li +4 authors

Zhiliang Wu, Jihao Gu, Kun Li, Hanyu Xuan, Hehe Fan, Meng Wang, Weisi Lin

First announced by arXiv: 2026-11-13

vision
Abstract from arxiv

Existing shadow removal methods either perform global shadow removal or rely on precise shadow masks to guide restoration. Such paradigms limit their ability to selectively remove the shadow cast by a user-specified object in practical applications. To this end, we introduce an object-aware controllable shadow removal task, which allows users to specify a target object through a simple prompt and removes only its cast shadow while preserving other content. However, object-shadow asymmetry makes it challenging to establish accurate correspondence between the specified object and its spatially separated cast shadow and to confine restoration to the target shadow region. To address these challenges, we propose FreeDeshadow, a unified controllable shadow removal framework that supports both visual and textual prompts. It consists of an object-aware generator and a shadow-aware local refiner. The former interprets the user prompt and produces a high-quality shadow-free proposal, while the latter establishes the correspondence between the object and its cast shadow and constrains the generated changes to the target shadow region. To support training and evaluation, we further construct PCSRD+, a large-scale dataset containing 154,571 quadruples at 1024$\times$1024 resolution and covering six types of user prompts. The visual prompts include a dot, a line, and an object mask, while the textual prompts describe the target object using attributes, spatial relations, or states. Extensive experiments demonstrate that FreeDeshadow accurately removes the shadow of the target object under both visual and textual prompts while effectively preserving non-target regions. It also outperforms existing methods in shadow removal performance and controllability. Project Page: https://wzlwlm13.github.io/FreeDeshadow

preprintarxiv

Cleaning an Agent's History Hurts When the World Still Holds the Error

Egor Pakhomov, Anurag Koul, Hongjun Liu +3 authors

Egor Pakhomov, Anurag Koul, Hongjun Liu, Bo Pang, Shafiq Joty, Erik Nijkamp

First announced by arXiv: 2026-11-13

languagelearning optimizationreasoning agents
Abstract from arxiv

Tool-using language-model agents accumulate histories of their own mistakes, and a system that compacts, repairs or hands off such a history decides what it says about those mistakes while their effects stay in the world. Prior work cleans the history at a fixed world, where showing a corrected step and saying it is done move together. We ask whether the same correct content helps or hurts depending on how the history states it and what the world holds. We freeze an agent mid-task with its exact world state, edit the history or the world, and let a receiver model continue. On AgentCE-Bench, after the agent commits a wrong value, we hold the world fixed and add the correct value by two natural edits: a truthful note beside the preserved error or the agent's own completed action. The note raises recovery, the share of remaining targets the receiver gets right, from 51% under a matched neutral note to 66%. The completed action lowers it below the unedited history, from 50% to 44%, so the value's presence alone does not predict the effect. On AppWorld, after a database write, we vary whether its receipt reports success or failure and whether its effect is present. A success receipt lowers whole-task success from 45% to 28% while the effect is absent and raises it from 38% to 49% once present. The two effects differ by 27 points pooled over two receivers, a difference positive in six of eight receivers at 97.5% and seven at 95%. A history edit's effect thus depends on the world it describes, and a history made to look clean while the world still holds the error can leave an agent worse off than no edit at all.

preprintarxiv

Closing the Loop: Co-Evolving EM for Irregular Time Series Generation in Lifted Representations

Idan Arbiv, Gal Fadlon, Omri Azencot

Idan Arbiv, Gal Fadlon, Omri Azencot

First announced by arXiv: 2026-11-13

learning optimization
Abstract from arxiv

Regular time-series generation from irregular time series is typically reduced to one-shot imputation: complete each partial sequence once, freeze the completions, and train a generator on the resulting surrogate dataset. We show that this open-loop strategy is structurally flawed. Point imputers collapse ambiguous missing regions to conditional averages, while stochastic imputers become stale as the generator evolves; the binding constraint is therefore iteration, not imputer design. We introduce a co-evolving Monte Carlo EM framework that closes the imputer-generator loop. Each E-step samples missing values from the posterior of the current generator, and each M-step retrains the generator on these completions, preserving unconditional generator training. Since our diffusion prior operates in a lifted 2D representation while observations live in time-series space, we further introduce Posterior Sampling for Lifted Representations (PSLR), which enforces operator-consistent conditioning, uncertainty preservation, and manifold consistency during the E-step. Across ten datasets and four corruption regimes, our method achieves state-of-the-art performance, improving over the strongest baselines by 59% in discriminative score, 15% in predictive score, 73% in Context-FID, and 71% in feature-correlation error. Ablations show that our approach reduces average discriminative score by 73% over single-step baselines, closing most of the remaining gap to the clean-data oracle. Code is available at https://github.com/azencot-group/PSLR.

preprintarxiv

Abstention Is Not Threshold Compliance: Stated Scoring Rules in Open Language Models

Dongsun Moon

Dongsun Moon

First announced by arXiv: 2026-11-13

languagereasoning agents
Abstract from arxiv

Kalai et al. propose stating the scoring rule in the prompt with an explicit threshold t, so that a model should answer only when its probability of being correct exceeds t. We probe each requirement of compliance on shared items, in 12 checkpoints from three post-training stages of four fully open 3-8B recipes, with exploratory models up to 32B. At t = 0.75, a line limiting the reply to an answer or "I abstain" (the abstention format) raised the mean abstention rate from 20% to 91%. Yet within this format, raising t from 0.5 to 0.9 changed it by only -0.8 percentage points. Without the line, the higher threshold even lowered abstention, and up to 40% of a checkpoint's items received decisions that no compliant policy with threshold-independent beliefs could make. Given the probability that their answer was correct, the checkpoints barely responded to the threshold in a forced-choice probe. Comparisons with self-assessment signals such as P(True) depended on the grader and on answer drift between prompts. An external threshold on recalibrated P(True) made coverage non-increasing in t by construction. Calibrated on the evaluation distribution (post hoc), it beat the abstention format at t = 0.5; calibrated on TriviaQA and applied to PopQA, however, it fell below that format at 0.75. On OLMo 3 7B, decisions trained on threshold-conditioned targets did not follow unseen thresholds consistently. A post-hoc variant trained at four thresholds answered less as t rose but still failed a pre-specified test, mainly on agreement with its training rule. A high abstention rate is not evidence of compliance: evaluations using stated scoring rules should report how often models abstain, whether decisions respond to the stated threshold, and which items they answer.

preprintarxiv

Galop: Billion-Scale Genetic Programming on a Single GPU

Xavier Hill Roy, Stephen Kelly

Xavier Hill Roy, Stephen Kelly

First announced by arXiv: 2026-11-13

learning optimization
Abstract from arxiv

Genetic programming (GP) evolves populations of heterogeneous programs, whereas GPUs achieve their highest throughput when groups of threads execute the same instructions. To better align GP evaluation with GPU execution, we present Galop, a hybrid CPU-GPU linear GP system that assigns one individual to each CUDA warp, a group of 32 threads that execute the same instructions across different fitness cases. We compare Galop against EvoGP, a recent high-performance GPU-based tree GP system, on data-rich symbolic regression across two unique GPUs, varying population size and fitness-case count. Galop completes evolutionary runs faster than EvoGP in 27 of the 28 main comparisons, with speedups of up to 13.6 times. Solution quality comparisons favour EvoGP overall, although Galop remains competitive on several problems. Galop also completes a full generation with 2.15 billion individuals on a single H100 GPU, approximately 4.0 times the largest population achieved by any other tested GP system. All five runs at this population complete 100 generations, with a median wall-clock time of 2.23 hours. These results make it practical to study GP populations in the billions on a single GPU. Galop's source code, experimental data, and reproducibility scripts are publicly available at https://github.com/xavierhillroy/Galop.

preprintarxiv

MITRA-Swarm: Injecting Explicit Linguistic Structure into Foundation Models for Classical Machine Translation

Kayshav Bhardwaj, Pragun Seth, Kush Bhardwaj +2 authors

Kayshav Bhardwaj, Pragun Seth, Kush Bhardwaj, Vijay Nagasamy, Kurt Keutzer

First announced by arXiv: 2026-11-13

foundation modelslanguagereasoning agents
Abstract from arxiv

Classical Sanskrit poses three challenges unresolved by single-call monolithic models: morphophonemic fusion Sandhi, dense nominal compounding Samāsa, and discourse-level zero-anaphora forced by poetic meter. We present MITRA-Swarm, a multi-agent architecture injecting explicit linguistic structure into an otherwise monolithic pipeline: a Sandhi-splitting agent, a parallel array of generalist gemma-4-26B-A4B-it and domain-specialized gemma-2-mitra-it translators, and a discourse-aware Ellipsis Restorer that consults corpus-derived register tags to avoid inventing active-voice subjects in impersonal legal and philosophical registers. Two separated judges evaluate quality: a routing judge ($J_{\text{route}}$, same model family as the generators) driving an internal critique loop, and an independent reporting judge ($J_{\text{report}}=$gpt-4o-mini, no shared vendor lineage) scoring every output exactly once. On a stratified benchmark of 1,000 verses across five registers (23,990 scored verdicts, 12-condition pre-registered ablation), the full swarm reaches $85.90$ MQM against a monolithic baseline of $85.27$ ($Δ=+0.63$, $p_{\text{adj}}=0.0018$, 95% CI $[+0.30,+0.97]$), and isolated discourse restoration alone accounts for $+0.44$ MQM ($p_{\text{adj}}=0.0288$), outperforming a call-count-matched self-refinement baseline by $+0.51$ MQM ($p_{\text{adj}}=0.0288$). Severe per-verse failures (MQM$<80$) fall from $9.0\%$ to $7.6\%$, at roughly $9\times$ the per-verse inference cost of the baseline. Additionally, $J_{\text{route}}$ scores its own family's outputs $2$-$7$ MQM higher than $J_{\text{report}}$ on every one of the 12 conditions, empirically motivating the two-judge split; and a preflight screen of a 192,151-entry retrieval index against all 12 evaluation works verifies 0 reference leaks across 269 runtime checks. These are small gains on a metric saturating near 85-86.

preprintarxiv

Fast Slice-Wise Deep Learning Motion Correction for Spinal Cord Functional MRI

Thomas Dagonneau, Sandrine Bédard, Nontharat Tucksinapinunchai +4 authors

Thomas Dagonneau, Sandrine Bédard, Nontharat Tucksinapinunchai, Uten Yarach, Loïc Le Folgoc, Pietro Gori, Julien Cohen-Adad

First announced by arXiv: 2026-11-13

vision
Abstract from arxiv

Respiratory and bulk motion displace the spinal cord between the volumes of a functional MRI time series. Current methods such as the Spinal Cord Toolbox (SCT) take several minutes per time series, and general-purpose brain tools register whole volumes, whereas the cord moves mainly slice by slice. We propose a lightweight convolutional neural network that regresses the in-plane translation aligning each axial slice to a reference. Two models, trained without motion labels, correct large and then residual motion. On a 3T (120 scans) and a 7T (59 scans) dataset, the proposed method improves on SCT on every metric (tSNR +10 to +25%, DVARS -20 to -38%), leaves almost no slice displaced by more than one voxel (<0.1%), performs on par with a slice-wise AFNI variant and better than slice-wise FSL FLIRT, and runs in about 10 s per scan, about 30 times faster than SCT. The code is available at https://github.com/ivadomed/moco-dl and the method is available in SCT v7.4.

preprintarxiv

Together Compiler: Continual Compilation for AI Inference

Reyna Abhyankar, Dylan Lim, Drew Wadsworth +9 authors

Reyna Abhyankar, Dylan Lim, Drew Wadsworth, Venkata Hasith Vattikuti, YuTong Julia Zhang, Jung Hwan Heo, Taeksang Kim, Kesavan Ramakrishnan, Evangelene Stanley, Xinyu Fang, Daniel Y. Fu, Simran Arora

First announced by arXiv: 2026-11-13

reasoning agents
Abstract from arxiv

Today's AI workloads are rapidly changing: new models are released more often, model architectures are increasingly diverse, and traffic distributions change frequently. The challenge we face at Together AI is continual optimization for these changing workloads. Compiling a deployment into an optimized forward pass is a challenge that is well-suited for agent hill-climbing. Existing systems struggle with continual compilation due to a tradeoff between measurement throughput and fidelity. Many systems use agents to optimize individual kernels from the inference forward pass using offline benchmarks, but timings are low-fidelity because of data-dependent behavior in primitives like mixture-of-experts (MoE). On the other hand, systems use agents to directly optimize end-to-end engines but suffer from slow engine launch times. To address these challenges, we build Together Compiler (TKCC), the first truly continual agent loop designed for AI inference optimization. To address the throughput-fidelity tradeoff, we propose: (1) a Tiered Persistent Engine that decouples engine launch overheads from forward pass computation; (2) a cost model to select the parallelism configurations that seed the optimization process. We also conduct extensive ablations on agent context management and present our insights on deploying TKCC in production. TKCC outperforms highly-optimized hand-written Megakernels on the Llama 1B workhorse workload by up to 1.7x and the best open source engines on the latest frontier models (Kimi K3, GLM 5.3, MiniMax M3.1 Flash, DeepSeek v4.1 Flash) across hardware platforms by up to 2.6x higher TPS on the InferenceX benchmark. TKCC also sets the state-of-the-art performance at batch size 1, with 1111 TPS on M3.1 Flash and 1205 TPS on DeepSeek v4.1 Flash. Finally, TKCC generalizes beyond frontier language model inference. TKCC serves trillions of tokens across our production fleet.

preprintarxiv

Trait Hitchhiking in LLM Post-Training

Adil Soubki, Elizabeth S. Z. Tan, Miles Cranmer

Adil Soubki, Elizabeth S. Z. Tan, Miles Cranmer

First announced by arXiv: 2026-11-13

languagelearning optimization
Abstract from arxiv

Reinforcement learning (RL) is now a standard stage of language model post-training, and the compute devoted to it is growing rapidly. While it has been shown to improve capabilities, particularly for tasks with verifiable rewards, it can also change behaviours and internal properties that the reward does not explicitly target. We present a formalism to describe the evolution of "traits" during RL post-training and, in doing so, recover the Price equation from evolutionary biology. We use this to develop a mathematical model for what we call "trait hitchhiking": the amplification or suppression of non-targeted traits as training redistributes probability across completions. We evaluate this Price update expression on language models trained with GRPO and show that the Price estimates closely track the sampled trait dynamics. These results identify trait hitchhiking as a mechanism by which RL post-training can silently amplify undesirable traits.

preprintarxiv

Terrain-Aware Thompson Sampling for Multi-Drone Search and Rescue

Riana Gagnon Souleiman, Giovanni Beltrame

Riana Gagnon Souleiman, Giovanni Beltrame

First announced by arXiv: 2026-11-13

robotics
Abstract from arxiv

In this work, we propose a decentralized terrain-aware multi-drone Thompson Sampling algorithm tailored for search and rescue (SAR) applications. A probability distribution of locating a lost person is informed by terrain data, using established SAR models (displacement ring, elevation, watershed, dispersion angle, and track offset) derived from historical SAR case data on lost person behavior. Exploration is conducted by a team of drones using Thompson Sampling, where candidate actions are selected by drawing samples from this probability distribution, represented by a per-cell Gaussian posterior whose mean and variance are updated through Bayesian inference. This approach balances exploration of uncertain regions with exploitation of high-probability areas to locate a lost person. Observation uncertainty is incorporated through fractional forest coverage, which measures the percentage of an observation occluded by tree coverage and accounts for reduced detection performance in densely vegetated areas. Simulations using real SAR case data (YOSAR) evaluate the proposed method against a coverage planner and a probabilistic active search method in search regions of 1.5 and 3.8 square kilometers, with 5, 10, 15, and 20 agents flying at 50 km/hr. We demonstrate that our method achieves a success rate of approximately 87% within 60 to 247 seconds in the 1.5 square kilometer region, and success rates of approximately 75% to 91% within 161 to 466 seconds in the 3.8 square kilometer region, across different agent configurations.

preprintarxiv

What Entropy-Based View Selection Cannot See: An Audit of Test-Time Adaptation for Vision-Language Models

Sivasuthan Sukumar, Uthayasanker Thayasivam

Sivasuthan Sukumar, Uthayasanker Thayasivam

First announced by arXiv: 2026-11-13

reasoning agentsvision
Abstract from arxiv

Vision-language models such as CLIP can classify images into categories they were never explicitly trained to recognize, but accuracy degrades when test images differ in style or source from their training data. A common remedy generates randomly cropped and flipped copies of each test image, called views, and combines their predictions. Since some views miss the object, methods must identify which views to trust. The dominant approach uses prediction entropy to measure confidence and retains the most certain views. Other methods use signals such as embedding typicality. Despite repeated refinements of these scoring functions, the information they rely on, the model's own response to each view, remains largely unexamined. We investigate this choice while holding other pipeline stages fixed. First, we show that view selection, rather than the gradient update, accounts for almost all accuracy gains. We then examine whether entropy identifies useful views when predictions are wrong, despite at least one view predicting the correct class. In these cases, entropy ranks correct views above incorrect ones no better than chance, recovering only about one-twentieth of the accuracy recovered by an oracle over the same views. This finding holds across two image encoders and under a control ruling out an obvious statistical artifact. An exact analysis explains this limitation: entropy measures confidence but not which class a view favors. We identify an unused signal already available in the augmentation pipeline: the fraction of the original image retained by each crop. Ranking views by this quantity improves accuracy on 14 of 15 benchmark datasets over entropy and on 11 of 15 over the strongest scoring function tested. A pre-method measurement also predicts which datasets benefit.

preprintarxiv

Beyond Row-Level Prediction: Evaluating and Learning Composition for Reusable Table Representations

Ali Younes, Saeed Ghoorchian, Maximilian Schambach +2 authors

Ali Younes, Saeed Ghoorchian, Maximilian Schambach, Georgia Chalvatzaki, Johannes Höhne

First announced by arXiv: 2026-11-13

learning optimization
Abstract from arxiv

Tables are compositional, with content spread across a schema, cell values, rows, and columns. Large-scale operations such as retrieval, overlap estimation, deduplication, and data curation need these local signals composed into one reusable embedding vector per table. Table representation methods, however, are mostly evaluated through row-level prediction or supervised downstream tasks, which do not measure table-level embedding quality directly. We introduce a benchmark that evaluates table-level embeddings through four desiderata: consistency under partial views, discriminability across label granularities, robustness to benign perturbations, and efficiency. Across twelve methods on synthetic and real corpora, lightweight schema- and text-based fingerprints are among the strongest. The sharpest contrast comes from two tabular foundation models, TabPFN and ConTextTab: although both are strong row-level predictors, their states, composed by mean, max, or three further parameter-free pooling rules, trail the strongest fingerprints. We therefore ask whether the gap lies in the encoder or in the compositional rule. Keeping ConTextTab frozen, we learn only the compositional rule over its cell states, using support-conditioned distribution matching (SCDM), a self-supervised loss we propose that matches the distribution of embedding similarities to the distribution of partition overlaps within and across tables. Across all fifteen head configurations, the learned rule raises the Spearman correlation between embedding similarity and partial-view overlap from 0.04 to 0.43-0.50 on synthetic data and improves triplet ranking, clustering, and retrieval, with smaller but systematic gains on real data. A downstream parent-table retrieval task confirms these findings. How local states are composed, and with what objective, substantially affects the quality of table-level embeddings.

preprintarxiv

Probe What-If Worlds Before You Ship: Persona Simulation for Real-World Decision Making

Julie Zhu, Jianheng Hou, Zhixu Silvia Tao +21 authors

Julie Zhu, Jianheng Hou, Zhixu Silvia Tao, Shi Bo, Yilan Fan, Yixuan He, Zhiyuan Hu, Shirley Huang, Fangyu Liu, Yijun Wang, Zibu Wei, Haimingyue Xu, Zhen Xu, Zichen Yuan, Jierui Zuo, Jintao Huang, Yifan Wang, Qianfeng Wen, Jiahan Li, Zhejian Peng, Bowen Yang, Shushu Liang, Yuexing Hao, Xiaomin Li

First announced by arXiv: 2026-11-13

language
Abstract from arxiv

User research and system testing reveal whether products and services meet user needs, yet recruiting specialized participants, designing controlled what-if comparisons, and repeating studies as designs evolve often bottleneck design iteration and data-driven decision-making. We present a framework and execution infrastructure that translates real-world industry inquiries into controlled experiments with simulated users. Our methodology formally defines the target cohort, experimental stimuli, and decision-relevant metrics, while the pipeline supports large-scale concurrent runs over cohort-stimulus-environment combinations (survey, web, chatbot, and OS app), auditable trial-level logging, cohort-level outcome aggregation, and cross-cohort and cross-variant statistical reports. Across twelve industry partner cases spanning five domains, we show how constructing hard-to-reach or what-if persona cohorts supports exploratory studies and distills simulated responses into testable design hypotheses. Cross-model comparisons on selected cases largely preserve primary conclusions, whereas perturbing or ablating persona attributes alters service outcomes, effect sizes, and inter-group differences, indicating that grounding target cohorts in detailed personas materially shapes simulation results. Together, these studies show that persona-enabled agents can test specified cohorts at design speed, diagnose systems under test, and deliver directional evidence that focuses problem definition, steers design and strategic decisions, and guides further study.

preprintarxiv

SyntheIce: A Synthetic Radargram Annotation Benchmark for Englacial Layer Reconstruction

Muhammad Behroze Hassan, Bayu Adhi Tama, Sanjay Purushotham +1 authors

Muhammad Behroze Hassan, Bayu Adhi Tama, Sanjay Purushotham, Vandana P. Janeja

First announced by arXiv: 2026-11-13

vision
Abstract from arxiv

Englacial layers in radio-echo-sounding radargrams (subsurface ice images) encode the climate and flow history of ice sheets and glaciers. However, large-scale extraction is hindered by time-consuming manual annotation and semi-automated detectors that produce fragmented rather than continuous horizons. We address two challenges: reconstructing complete layers from partial detections across radargram profiles and validating layer continuity across adjacent tile boundaries. We introduce a slope-aware greedy fragment-linking algorithm that reconstructs full-width layers using row-similarity gating and an exponentially weighted moving average (EWMA) slope predictor. Each reconstruction receives a content-based hash ID for stable cross-radargram identification. For quantitative evaluation, we generate SyntheIce, a synthetic dataset of 600 radargrams spanning multiple complexity levels and realistic noise conditions, including 150 wide radargrams divided into four adjacent tiles. All samples include exact ground-truth annotations. We evaluate the dataset using XCheck, a consistency-based validation framework, extending its original two-image setting to four consecutive tiles and enforcing continuity at every boundary. This demonstrates the dataset's utility for developing and evaluating real-world annotation algorithms. Our findings extend beyond ice science: when spatial features are fragmented but retain consistent local geometry, trajectory-based linking is more reliable than appearance-based matching. Evaluation against expert-labeled real radargrams demonstrates high precision, recall, F1-score, and mean IoU, supporting the method's transfer to real-world data.

preprintarxiv

SunLightBots: Turning sunlight into self-sustained, high-power motion in soft robots

Yoo Jin Lee, Melanie Torres Crespo, Zhongxiao Deng +4 authors

Yoo Jin Lee, Melanie Torres Crespo, Zhongxiao Deng, Andrew Tejada-Vega, Rebecca Kramer-Bottiglio, Wenzhong Yan, Yimin Luo

First announced by arXiv: 2026-11-13

robotics
Abstract from arxiv

Soft robots built from active materials can directly convert environmental stimuli into motion, offering a route to energy autonomy and embodied control. However, environmental energy is generally harvested at low power, whereas rapid, mechanically demanding actions require high mechanical output power. Reconciling this mismatch while sustaining self-regulated cyclic actuation is challenging, particularly at the subgram scale, where limited payload capacity constrains the integration of conventional power and control hardware. Here, we present SunLightBots, a class of subgram, sunlight-powered, untethered soft robots that achieve autonomous locomotion under continuous illumination near 1 Sun (100 mW/cm2). Their operation is enabled by a self-resetting snapping (SRS) mechanism that integrates liquid crystal elastomers (LCEs) with snap-through instability to accumulate solar energy supplied at low power, release it as rapid, high-power motion, and reset through photothermal-mechanical feedback. Through coordinated optimization of the LCE actuation temperature, photothermal conversion, actuation stress, and structural energy barrier, the system operates near 1 Sun and achieves a peak mass-specific mechanical power of 89 W/kg. At this irradiance, the optimized LCE generates an actuation stress of 0.1 MPa at approximately 60 deg C. SunLightBots demonstrate sustained swimming, crawling, and repeated impulsive jumping. Under 1.5 Suns, the swimmer advances more than three body lengths per cycle. This work provides a materials-mechanics route toward autonomous soft robots powered entirely by the environment.

preprintarxiv

When Should Semantic Priors Enter Partially-Identified Causal Discovery?

Muhammad Ahmed Mohsin, Saad Bin Ather, Valentin Six +3 authors

Muhammad Ahmed Mohsin, Saad Bin Ather, Valentin Six, Lancelot Da Costa, Tommaso Salvatori, Bernhard Scholkopf

First announced by arXiv: 2026-11-13

reasoning agents
Abstract from arxiv

Observational causal discovery often leaves causal di- rections unresolved or unstable. Large language mod- els (LLMs) can provide semantic information to resolve these directions, but incorrect proposals can exclude the true causal graph and invalidate downstream inference. We study three separate decisions: where to allocate semantic queries, which proposals to admit, and which causal graphs to retain. Our framework distinguishes population non-identifiability from finite-sample insta- bility, evaluates semantic orientation against statistical and random controls, and examines calibrated admis- sion and error-tolerant DAG retention. Experiments across benchmark networks show that semantic infor- mation improves recovery over coin-flip orientations, but targeting unstable edges has no established ad- vantage over matched-random allocation. The tested admission certificate cannot support nontrivial admis- sion with the available calibration tasks. Moreover, downstream retention depends on reference adequacy: learned references contain the true DAG in only three of 80 planned tasks. These findings separate the value of semantic information from the reliability of its alloca- tion, admission, and retention, and identify calibration size and reference adequacy as key limitations.

preprintarxiv

UniPipe: Unified Inference Serving and Fine-Tuning for Deadline-Constrained DNN Surrogate Workflows

Jaiaid Mobin, Avinash Maurya, M. Mustafa Rafique +1 authors

Jaiaid Mobin, Avinash Maurya, M. Mustafa Rafique, Bogdan Nicolae

First announced by arXiv: 2026-11-13

learning optimization
Abstract from arxiv

DNN surrogates are increasingly used to accelerate scientific workflows that must process high-rate sensor streams in near real time, exemplified by Ptychographic imaging at next-generation X-ray light sources. Such surrogates deliver 100-1000x higher throughput than iterative solvers, enabling faster steering of experiments. To preserve surrogate accuracy during an experiment, as the input distribution evolves, state-of-the-art deployments freeze the pretrained surrogate or run a separate CUDA context that fine-tunes concurrently with serving. However, multi-context concurrency causes significant contention on the GPU, inflating both the mean and the variance of inference latency, triggering deadline misses in workflows where sensor data has to be processed within a hard per-unit deadline. To address this challenge, we propose UniPipe, a single-context system that unifies inference serving and fine-tuning of a DNN surrogate on a single edge GPU. UniPipe is driven by three key insights: both serving and fine-tuning require a forward pass and can be fused into one; the per-sample latency of modern GPUs is non-linear in batch size; and the workload is repetitive across intervals, enabling offline scheduling. Specifically, UniPipe fuses training and inference data in a single forward pass with a selective backward pass on the training subset, and uses a polynomial-time scheduler that exploits the latency curve to compute the batch composition. We evaluate UniPipe on a real-world Ptychographic reconstruction workflow with sensor rates up to 5000 patterns/sec.; UniPipe lowers reconstruction MSE by up to 30% and inference MSE by up to 50% over a pretrained baseline, and reduces the deadline-miss rate by 13%-22% over a multi-context baseline. UniPipe's unified deployment reduces the latency spread from 106% to 57% of the mean, and increases total throughput compared to multi-context baseline.

preprintarxiv

From Nodes to Pixels: Topological and Structural Two-View Graph Imaging

Md Joshem Uddin, Soham Changani, Sai Karthik Navuluru +2 authors

Md Joshem Uddin, Soham Changani, Sai Karthik Navuluru, Cuneyt Gurcan Akcora, Baris Coskunuzer

First announced by arXiv: 2026-11-13

learning optimizationreasoning agents
Abstract from arxiv

Graph learning still lacks a broadly reusable input interface comparable to patches in vision or tokens in language. Unlike images or text, graphs are irregular, vary in size and density, and are invariant under many equivalent node labelings, making it difficult to standardize inputs across tasks and architectures. We introduce G2Image, a fixed-budget graph imaging framework that converts each graph into two complementary image-like views. TopoGrid provides a stable topological view by encoding an intrinsic bifiltration as a compact multipersistence surface grid, while GraphGrid provides a higher-bandwidth structural view by aggregating edge densities between groups defined by intrinsic node scores. Both views are permutation-invariant and come with resolution-controlled stability guarantees. Because the resulting representations are fixed-size tensors, they can be processed by lightweight 2D encoders and aligned through a supervised cross-view contrastive objective. Across graph classification and molecular prediction benchmarks, G2Image achieves strong average performance among the compared methods, improves over single-view and standard fusion variants, and extends to attribute-rich molecular graphs without changing the overall architecture. These results suggest that fixed-budget graph images offer a practical and reusable interface for graph representation learning.

preprintarxiv

Benchmarking Multilingual Document Understanding in Vision-Language Models

Matteo Gabburo, Daniela Baiamonte, Stefano Simonazzi +3 authors

Matteo Gabburo, Daniela Baiamonte, Stefano Simonazzi, Silvia Gardini, Sean Cesare Lazzeri, Andrea Zugarini

First announced by arXiv: 2026-11-13

languagereasoning agents
Abstract from arxiv

Recent Vision-Language Models (VLMs) perform strongly on multimodal tasks, Document Understanding (DU) among them, but the benchmarks that measure them mostly cover English documents and narrow problems. How well current VLMs generalize across languages, tasks, styles and layouts therefore remains unclear. We introduce MDU, a multilingual DU benchmark of 23146 examples spanning 13 tasks and five languages: English, French, German, Italian, and Spanish. The tasks isolate complementary capabilities, from surface-level text and Markdown extraction through summarization, classification and named-entity recognition to layout-sensitive passage and layout extraction. We evaluate 45 state-of-the-art open-weight VLMs from 1.7B to 235B parameters, plus one commercial system under the same protocol. The benchmark is far from saturated: the best average is 69.0, only eight of the 45 evaluated systems reach 65, and the two best results come from 27B models rather than from the largest systems we evaluate. Performance is uneven in ways that scale does not fix. The gap between the strongest and weakest language does not close as models grow, and it survives controlling for the domain composition of each language. Capabilities also do not improve together: once we partial out overall model quality, per-task scores are largely uncorrelated, and transcription ability in particular trades off against semantic extraction.

preprintarxiv

Graph Label Alignment: A Diagnostic Atlas for Graph Classification

Neelam Akula, Varun Shiralkar, Murat Kantarcioglu +1 authors

Neelam Akula, Varun Shiralkar, Murat Kantarcioglu, Baris Coskunuzer

First announced by arXiv: 2026-11-13

learning optimizationreasoning agents
Abstract from arxiv

Graph classification benchmarks vary widely in difficulty: some are solved by simple pooled baselines, while others remain challenging for GNNs and graph transformers. We take a data-centric view and ask whether benchmark success is driven by genuine structure--feature reasoning or by label signal already accessible from restricted marginal views. We introduce graph label alignment, a lightweight diagnostic that measures label accessibility through two supervised probes computed without training a graph encoder: structural label alignment (SLA), based on permutation-invariant topological and spectral summaries, and feature label alignment (FLA), based on pooled node features with edges removed. Across real graph classification benchmarks and multiple pooled backbones, the resulting SLA--FLA plane strongly organizes test accuracy, with low-alignment datasets forming a consistent regime where standard pooled pipelines struggle. To probe what lies beyond marginal accessibility, we introduce controlled interaction-dominated stress tests, including heterophilic ego-graphs and synthetic Twins tasks whose structural and feature marginals are matched across classes but whose role--feature couplings differ. These results show that high alignment explains much of easy benchmark performance, while low alignment reveals regimes where success requires preserving structure--feature correspondence beyond standard pooled representations.

preprintarxiv

Open-Source Transtibial Prosthetic Bypass Preserving Joint Alignment: Gait Learning and Kinematic Evaluation

Apolline de Vaulchier, Pablo Nocito, Yann Corbel +1 authors

Apolline de Vaulchier, Pablo Nocito, Yann Corbel, Nathanaël Jarrassé

First announced by arXiv: 2026-11-13

robotics
Abstract from arxiv

Prosthetic bypasses allow able-bodied individuals to walk with a prosthesis, enabling safe early-stage testing of new prosthetic devices before clinical trials. However, existing transtibial bypasses impose constraints, such as a contralateral shoe lift or a lateral offset of the prosthesis, and their effects on gait remain poorly documented. We present an open-source transtibial prosthetic bypass that preserves part of the sagittal knee mobility without these constraints. Six able-bodied participants familiarized themselves with the bypass for approximately 10~min, then walked on a treadmill with and without it, and joint kinematics were recorded with motion capture and then extracted with OpenSim. All participants were able to walk without handrail support, although single stance remained shorter on the bypass side (27% vs. 45% of the gait cycle). Bypass-side ankle kinematics were consistent with those reported for transtibial prosthesis users, whereas peak knee flexion during swing decreased from 61° to 36° (p<0.001) and hip range of the sound leg increased from 35° to 51° (p<0.01). Although gait with the bypass does not fully replicate that of prosthesis users, it yielded consistent gait patterns after only 10~min of learning. This bypass therefore provides a useful tool for the lower-limb prosthetics research community, facilitating preliminary experiments with able-bodied participants.

preprintarxiv

RECALL: Real-Time Recalibration of Multi-Camera Rigs via Learned OEM-Anchored Residuals

Mario Alberto Tapia de la Cruz, Charalambos Poullis

Mario Alberto Tapia de la Cruz, Charalambos Poullis

First announced by arXiv: 2026-11-13

vision
Abstract from arxiv

Multi-camera rigs ship pre-calibrated from the manufacturer, but their calibration goes stale each time the rig is repositioned, and drifts continuously through wear, thermal cycling, and vibration. Recovering it on-the-fly is hard: classical bundle adjustment needs a fresh multi-view capture and a global optimization whose cost grows with the rig, ruling out real-time use. In this paper we present RECALL, a feed-forward method that recalibrates a movable multi-camera rig in a single pass. From the nominal OEM calibration as an anchor and a brief observation of a known object, it predicts each camera's current intrinsics and extrinsics, written as a shared global rotation of the rig composed with a small per-camera drift. Instead of regressing the calibration in one shot, RECALL refines its estimate with a learned, rig-coupled recurrent operator that reprojects through a differentiable projection and applies attention across the cameras at each step, and solves for the shared rotational gauge in-network with a learned Gauss-Newton step rather than regressing it. Because the ground truth is synthesizable, we train entirely on data generated on-the-fly and need no captured dataset. On a synthetic 10-camera benchmark, RECALL is more accurate than per-camera and rig-coupled joint bundle adjustment ($0.26^\circ$ vs $0.47^\circ$ rotation, $5.4$ vs $8.9$ px focal) and runs about $10^5\times$ faster at about $0.12$ ms per rig. On real data (TotalCapture, CMU Panoptic) it matches or beats offline rig-coupled bundle adjustment on absolute and relative rotation while running four-to-five orders of magnitude faster, and on the Panoptic surround rig it also matches the strongest image-only foundation models (VGGT, VGGT-$Ω$, $π^3$) on relative pose while staying metric and about $15\times$ more accurate in focal length. Code: https://github.com/theICTlab/RECALL-multicam

preprintarxiv

SynthCorp: A Platform for Creating Benchmarks for Data Agents

Jiani Zhang, Manav Garg, Eugenie Lai +5 authors

Jiani Zhang, Manav Garg, Eugenie Lai, Cheng Zhen, Eva Sharma, Yang Pan, Yaming Lin, Alon Halevy

First announced by arXiv: 2026-11-13

reasoning agents
Abstract from arxiv

Current benchmarks for LLM-driven data agents saturate quickly and cannot target specific capabilities. We present SynthCorp, a platform that creates customizable data estates and benchmarks by simulating how business events produce and modify corporate data over time. Each data estate combines relational databases, a document store, and a data catalog, paired with evaluation tasks and executable gold programs. First, given a scenario and target complexity dimensions, SynthCorp synthesizes reusable event types and compiles each into a program over predefined simulation primitives. Second, SynthCorp schedules event instances on a virtual calendar, interleaving routine and disruption events so that tables, documents, and catalog entries remain consistent. Third, SynthCorp executes the calendar to materialize the data estate and records every operation in a simulation log. Fourth, SynthCorp generates analytical tasks and derives verified gold programs from the log. We use SynthCorp to generate initial evaluation benchmarks comprising seven data estates (up to 1K operational tables) and 256 tasks across five domains. The data estates pass deterministic pipeline checks (such as foreign-key violations, undeclared writes, recurrence violations, and database execution failures) with a near-zero error rate, introduce cross-asset complexities, and cover all requested dimensions. Evaluating three models across four data-context settings shows that the benchmarks separate data contexts and model tiers. Auditing candidate queries against causal provenance in the simulation log verifies gold programs without manual annotation, including on multi-event composite tasks that expose reasoning failures in state-of-the-art agents. We hope the community uses SynthCorp to generate more challenging and useful data estates and benchmarks that foster the development of enterprise data agents.

preprintarxiv

Rational Use of Time for Social Coordination in Humans and AI

Xiulin Yang, Sirisha Gudavalli, Jennifer Hu

Xiulin Yang, Sirisha Gudavalli, Jennifer Hu

First announced by arXiv: 2026-11-13

languagereasoning agents
Abstract from arxiv

All social interactions are situated in time. Previous studies have shown that humans are able to use implicit temporal cues, such as the duration of a pause, to infer others' mental states. But how do humans use temporal cues for dynamic, fine-grained social coordination? And can AI systems do the same, without ``experiencing'' continuous real-time as humans do? We study these questions using a cooperative card game task called The Mind, which requires players to achieve a shared goal across repeated interactions but forbids explicit communication: the only available signal is the passage of time. We evaluate 150 humans, and two classes of AI models which reflect different ways of operationalizing time in artificial systems: computer-use agents (CUAs), and standard text-prompted LLMs. We find that human behavior is best captured by a Bayesian strategy that infers a partner's belief from their timing---and similarly for frontier models. However, unlike humans, none of the tested models seem to efficiently adapt their behaviors to their partners' over the course of the game. From a cognitive perspective, our findings suggest that humans perform implicit probabilistic reasoning about continuous-time signals in dynamic social settings, and that such reasoning mechanisms may emerge without real-time experience. Practically, our study highlights new challenges for AI systems to coordinate efficiently with other agents by learning to interpret and adapt to their implicit signals in dynamic settings.

preprintarxiv

Leaderboard Hacking: Preference-Based Model Evaluations are Vulnerable to Manipulation

Eve Fleisig, Joachim Baumann, Dylan Hadfield-Menell +1 authors

Eve Fleisig, Joachim Baumann, Dylan Hadfield-Menell, Dirk Hovy

First announced by arXiv: 2026-11-13

reasoning agents
Abstract from arxiv

Arena-style rankings of language models are a widely used evaluation framework in online leaderboards, research papers, and model evaluation in production. These rankings aim to realistically assess model performance using pairwise comparisons of anonymous models on user prompts, aggregating votes via the Bradley-Terry model to produce the final ranking. However, we find that these leaderboards are highly vulnerable to several manipulation strategies that unscrupulous providers could use to boost a model's rank: strategic voting (individual votes that boost a model's rank, including votes on irrelevant models), strategic prompting (choosing prompts that favor a model), and strategic nomination (boosting the rank of a model by inserting irrelevant models). We demonstrate these effects on public leaderboard data, analyze the circumstances that cause these vulnerabilities, and describe approaches to safeguard against each attack. Notably, these issues hold for any preference-based evaluation using the Bradley-Terry model that allows open prompting, voting, or model submission. These manipulations boost the rank of nearly every tested model; each attack can boost a model by up to 15-20 places in a ranking and, combined, they can boost a model by up to 45 places.

preprintarxiv

An LLM Is All You Need: Data-Efficient Adaptation to Full-Duplex, Word-Level Spoken Dialogue

Tairan Chen, Richard Yucheng He, Yihang Liu +2 authors

Tairan Chen, Richard Yucheng He, Yihang Liu, Baodong Cao, Chen Xu

First announced by arXiv: 2026-11-13

language
Abstract from arxiv

Full-duplex spoken dialogue, in which a system listens while it speaks and can backchannel, yield the floor, and be interrupted, is usually obtained by training on hundreds of thousands to millions of hours of speech. This is costly, rewrites the language model that supplies the system's knowledge, and still leaves turn-taking either implicit or decided coarsely rather than word by word. We show that a frozen, general-purpose multimodal LLM can instead be adapted into a word-level full-duplex system with little data. Our system is a thinker-talker cascade on an 80 ms frame clock. The thinker keeps the LLM frozen and trains only lightweight adapters and a few new token embeddings (97M parameters); at every frame it emits one token of agent text, with pauses, yields, and onsets represented as tokens of the same vocabulary and kept legal by a finite-state machine. A four-stage curriculum, each stage anchored to the model's own behavior under the previous format, performs the conversion in about 30 GPU-hours. The talker reads only the thinker's discrete tokens, so the two halves are trained independently. On Full-Duplex-Bench and InteractionBench, the system is competitive with open-source full-duplex systems built through large-scale speech training: among them, it is the least likely to take the floor when the user merely pauses, backchannels most often, and leads in spoken-task accuracy, groundedness, and coherence over long conversations.

preprintarxiv

DeNA-SGN: Deterministic Neighborhood Aggregation for Verified Spiking Graph Inference

Sanjay Das, Xi Zhang, Emily Herron +1 authors

Sanjay Das, Xi Zhang, Emily Herron, Prasanna Date

First announced by arXiv: 2026-11-13

learning optimization
Abstract from arxiv

Graph learning on neuromorphic substrates requires an explicit spiking realization of neighborhood aggregation, supporting control and classification. We propose DeNA-SGN, which combines fixed OR and tie-inclusive Majority aggregation of binary node features with a learned signed-weight threshold classifier. A degree-dependent inhibitory construction implements Majority through shared control neurons, enabling a complete circuit compilation into SuperNeuroMAT. Across seven datasets and 41 mapped classifiers, native simulation matches the exported ANN on every evaluated class decision and output-spike pattern. On the new confirmatory Amazon Photo dataset, DeNA-SGN achieves 83.93+-0.93% current-readout accuracy and 71.61+-2.22% strict output-spike accuracy; Wiki-CS reaches 75.99+-0.54% over 20 public masks. In the matched Cora study, we achieve 74.90+-2.67% on current-readout and 68.57+-2.94% strict spike accuracy. Complete circuit and activity accounting supports analytical energy comparisons and identifies the cost of clocking inactive neurons in large deployments. Current argmax is decoded on the host, and energy is modeled rather than measured on hardware. These findings establish a reproducible route to useful graph classification with decision-matched spiking execution and explicit implementation costs.

preprintarxiv

RE-PRESENT: Evidence Triangulation across Alternative Candidate-Space Views for Efficient LLM Ensemble Inference

Juming Xiong, Weixin Liu, Kevin Guo +7 authors

Juming Xiong, Weixin Liu, Kevin Guo, Congning Ni, Chongyu Qu, Junchao Zhu, Katherine Brown, Chao Yan, Bradley Malin, Zhijun Yin

First announced by arXiv: 2026-11-13

reasoning agents
Abstract from arxiv

Heterogeneous large language models (LLMs) exhibit complementary strengths, motivating routing and multi-model inference methods that seek to identify reliable predictions for each input. However, exploiting this complementarity remains challenging even when a correct prediction is available within the ensemble. We investigate this challenge across heterogeneous LLM ensembles and find that a simple single-round aggregation rule is highly competitive with learned routing methods, yet substantial oracle headroom remains. Further analysis shows that the evaluated single-round response signals provide limited discriminative power for identifying reliable predictions. In contrast, controlled re-presentations of the same question elicit complementary evidence about prediction reliability. Motivated by these findings, we introduce RE-PRESENT, a training-free framework that selectively re-presents disputed questions and directly aggregates candidate-level evidence across alternative views. Across seven multiple-choice benchmarks and two heterogeneous model ensembles, RE-PRESENT improves over strong single-round aggregation while offering a favorable accuracy--cost trade-off compared with self-consistency and multi-model debate.

preprintarxiv

Robust Task-Aware State Estimation from Pre-trained Perception

Guy Azran, Sarah Keren

Guy Azran, Sarah Keren

First announced by arXiv: 2026-11-13

reasoning agents
Abstract from arxiv

Agents that perform state estimation using pretrained perception models, e.g., Vision-Language Models (VLMs), face the fundamental challenge that predictions are uncertain. Still, most approaches determinize these predictions into a single best-guess state and make decisions as if it were ground truth. This is brittle because when the guess is wrong, the agent's decisions may be unsafe or infeasible. At the other extreme, maintaining a full probability distribution over symbolic states is often intractable. We propose a middle path, drawing on the underlying probabilities the perception model already produces (e.g., a VLM's output-layer token distributions) to construct a $θ$-confidence Minimal Belief Set (MBS$_θ$), a minimal subset of states whose cumulative probability meets a user-specified belief-mass threshold. The minimality of an MBS$_θ$ ensures that the agent considers only the most likely states, enhancing robustness in downstream decision-making while maintaining computational tractability. Using a novel three-layered belief system, we efficiently approximate individual state probabilities from per-feature predictions and task-level constraints. On this belief construction, we base our MBS$_θ$ construction, while maintaining only a lean factored representation of the belief. Our algorithm, Probability-Ranked Valid State Search (PRaVaSS), takes advantage of the unique structure of this belief system to construct an MBS$_θ$ with correctness guarantees and polynomial complexity in the unconstrained case. Using a VLM as the backbone of our state estimator, we validate our method's theoretical properties, scalability with problem size, and improvement of downstream decision-making when used in a task-planning pipeline.

preprintarxiv

When Is Proof-Step Verification Worth Its Cost? Operating Requirements of Certify-or-Decline Verification for Open-Weight Mathematical Reasoning

Abhinav Gupta, Ben Slivinski, Yashica Patodia +1 authors

Abhinav Gupta, Ben Slivinski, Yashica Patodia, Michael Saldivar

First announced by arXiv: 2026-11-13

languagelearning optimizationreasoning agents
Abstract from arxiv

Proof-step verification rewrites a solver's answer into typed steps and uses LLM judges to certify or decline. We evaluate this pipeline on open-weight models across GPQA, AIME-2025, and a MATH-500 subset. Certification raises answer-key-match precision by 31 percentage points on the strong GPQA configuration and 46 on AIME, while answering only 35% and 21% of attempts. On near-ceiling MATH-500, gains are at most 5 points at a loss of 13-27 points of coverage. An output-token-matched self-consistency selector obtains 82.9% precision at 35% coverage, compared with the verifier's pooled 87.6% (paired difference: +4.8 points, 95% interval [-8.2, 18.5]). A diagnostic sample reveals a distinct failure mode: 52 of 88 completed problems end in invalid proof construction without ever reaching a judge. Finally, external-LLM acceptance of rendered certified chains is 61.0-67.6%, compared with 87.6% answer-key match. These audits lack human or formal-kernel ground truth. Open-weight auditor pairs show near-zero chance-corrected agreement, which measures consistency rather than accuracy against proof truth. These results motivate separate reporting of execution failures, selected-answer precision, and independently assessed proof validity. We provide per-configuration results and evaluation details in the supplement.

preprintarxiv

Distributionally Robust BEST-RQ For Speech Representation Learning

Muhammad Khan, Xiaodong Cui, Brian Kingsbury +1 authors

Muhammad Khan, Xiaodong Cui, Brian Kingsbury, George Saon

First announced by arXiv: 2026-11-13

language
Abstract from arxiv

BEST-RQ, based on random projection and quantization, has been a well-established self-supervised learning approach to learn speech representations from large-scale unlabeled audio due to its high performance and efficiency. However, in BEST-RQ representations learned by random quantization can be sensitive to the choice of quantization targets. To address this issue, we propose a distributionally robust extension of BEST-RQ that seeks representations robust across multiple quantization configurations. Specifically, we define groups by codebooks and formulate the training as a group Distributionally Robust Optimization (DRO) problem. Rather than optimizing solely for the average loss across groups, our approach adaptively reweights groups to emphasize those performing poorly. Additionally, we regularize DRO with a Kullback-Leibler divergence to prevent excessive distribution deviation. We evaluate our approach on the Multilingual LibriSpeech dataset and show that the distributionally robust BEST-RQ achieves superior performance over the conventional uniform average across multiple codebooks.

preprintarxiv

GenPFN: An In-Context Tabular Foundation Model Approach for Augmenting Small Datasets in Engineering

Noah J. Bagazinski, Cyril Picard, Faez Ahmed

Noah J. Bagazinski, Cyril Picard, Faez Ahmed

First announced by arXiv: 2026-11-13

foundation modelsreasoning agents
Abstract from arxiv

Costly simulations and experiments limit the data available for engineering design. This paper proposes GenPFN, a training-free candidate generator guided by a pretrained tabular Prior-Data Fitted Network (TabPFN) classifier with fixed weights. The context contains evaluated designs labeled desired (Class 1) or undesired (Class 0). GenPFN interpolates between desired designs and refines candidates using gated classifier gradients and Gaussian noise. Each candidate is evaluated using the problem's constraints and objective functions. Evaluated desired designs and their objective values augment the context for regression training. Evaluation spans fourteen selected problems, including a two-parameter benchmark and a 44-parameter hull problem. In paired comparisons, GenPFN with TabPFN v1 achieves higher class satisfaction and lower realism distance than the tested adversarial, variational-autoencoder and diffusion generators. The variational-autoencoder generator provides greater coverage and diversity. Relative to the synthetic minority over-sampling technique (SMOTE), GenPFN provides lower coverage distance and greater diversity. Among observed paired regression fits, augmentation reduces estimated AutoGluon validation mean squared error by 62.1% relative to context-only training. A hull-stern case study yields 99.3% propeller-compatible candidates versus 1.4% among randomly sampled context designs. Generation followed by evaluation produces sample sets that augment limited datasets and support training more accurate predictive models.

preprintarxiv

BirdDreams: Joint Perspective and BEV Video Generation for a Plausible, Reactive Driving World Model

Junyoung Seo, Jaewoo Seo, Seungryong Kim +3 authors

Junyoung Seo, Jaewoo Seo, Seungryong Kim, Laura Leal-Taixé, Xiuming Zhang, Seung Wook Kim

First announced by arXiv: 2026-11-13

vision
Abstract from arxiv

Driving world models aim to simulate how a scene unfolds under different ego actions. This requires plausible motion of surrounding vehicles as well as precise ego-trajectory control. Learning vehicle motion from perspective video alone is challenging because motion is entangled with appearance and vehicles can be occluded. Prior work therefore conditions video generation on layouts that specify vehicle motion, providing control over surrounding vehicles but requiring their future trajectories as inputs. We introduce BirdDreams, an autoregressive driving world model that jointly generates perspective video and a schematic bird's-eye-view (BEV) video, making scene layout part of the predicted future. The BEV, rendered from scene annotations for training, depicts roads and vehicles as simple shapes at a fixed metric scale, providing a top-down view of their positions and motion with reduced occlusion between vehicles. Perspective video, by contrast, retains scene details that the BEV omits, such as traffic signals and object appearance. Joint attention between two video diffusion transformers, one per stream, lets each stream inform the other during generation. The predicted BEV also provides a geometric interface for ego control: we translate and rotate its latest frame along an input ego trajectory and use the resulting frames as a soft condition that shows how the scene would appear to move if only the ego vehicle moved. At inference, BirdDreams generates both videos from a single camera image without requiring an HD map or vehicle-state inputs. On PhysicalAI-AV and nuScenes, BirdDreams follows input ego trajectories more faithfully than prior driving world models and generates more plausible ego motion without ego-action conditioning. Human evaluation shows that BirdDreams outperforms a perspective-only baseline in generating realistic surrounding-vehicle behavior.

preprintarxiv

What Do Tokenizer Comparisons Measure? Disentangling Granularity, Root Assignment, and Word-Level Readout across Amharic and Hausa

Abdul Rehman Antall, Baidya Nath Saha

Abdul Rehman Antall, Baidya Nath Saha

First announced by arXiv: 2026-11-13

languagelearning optimization
Abstract from arxiv

Morphology-aware tokenizers are expected to help morphologically rich languages, and their reported gains over byte-pair encoding (BPE) are usually credited to their linguistic units. A comparison with BPE cannot support this attribution, since such a tokenizer also segments text more finely (granularity) and, on token-level tasks, changes how much the sub-token readout matters. We separate both factors from the effect of grouping words by their roots (root assignment) in Amharic, a Semitic language whose Ge'ez abugida hides consonantal roots inside syllabic glyphs, with Latin-script Hausa as an orthographic contrast. Our design combines a lossless root-and-pattern tokenizer with controls that change one factor at a time, in two independent implementations. In both, real root assignment improves Amharic named-entity recognition over scrambled roots (+4.9 and +4.7 F1), and root-frequency structure explains little of this gain. In small from-scratch models, the readout choice moves F1 by up to 13.4 points and reverses the sign of every comparison of the root-and-pattern arms with BPE, whereas in two pretrained encoders it moves F1 by less than one point. The granularity effect changes sign across languages and implementations. We recommend decomposing tokenizer effects, and reporting the readout, before a gain is attributed to linguistic structure.

preprintarxiv

From Classification to Generation: Vision-Language Models for Roman Imperial Coin Description

Pedro Lourenco, Ognjen Arandjelović

Pedro Lourenco, Ognjen Arandjelović

First announced by arXiv: 2026-11-13

vision
Abstract from arxiv

Cataloguing a Roman Imperial coin involves reading its legends and identifying its imagery in order to match it to an entry in the standard reference catalogue. Previous work has automated individual parts of this process, such as identifying the issuer or recognizing reverse motifs, by formulating each part as a classification task over a predefined set of labels. We instead formulate coin description as image-to-text generation and adapt an open-weight vision-language model (VLM), Qwen3-VL-8B-Instruct, to four tasks: transcribing the obverse and reverse legends, and describing the obverse and reverse imagery. To obtain training data, we convert approximately 30,000 auction records into structured image-text pairs. We separate the two sides of each coin, extract and normalize legends and descriptions, and standardize obverse descriptions against a bust-type taxonomy. We then train a separate low-rank adapter for each task on a single 20GB GPU, and evaluate each task with metrics that we define to capture the key components of numismatic descriptions. The adapted models outperform both the unadapted base model and GPT-4.1 on all four tasks, reducing the normalized edit distance of legend transcriptions by over 40% relative to GPT-4.1. The improvement is smaller, but still present, on legends absent from the training data.

preprintarxiv

Discharge Summary Generation Is Not Just Summarization: A Reality Check for Evaluation

Radhika Dua, Frederick Wolff, Kyunghyun Cho +1 authors

Radhika Dua, Frederick Wolff, Kyunghyun Cho, Eric Karl Oermann

First announced by arXiv: 2026-11-13

languagelearning optimizationreasoning agents
Abstract from arxiv

Discharge summaries combine a retrospective account of hospitalization with discharge-time decisions. Physician-written references therefore contain both facts supported by prior clinical notes and reference-only content, potentially reflecting clinical synthesis, handoffs, or discharge planning. Reference-only evaluation can conflate failure to recover available information with failure to reproduce content absent from the model input. We introduce FairSumEval, an evidence-aware, fact-level framework that assesses whether each reference and generated fact is supported by the model's evidence and captured by the other summary. The resulting outcome map separates matches, omissions, and additions. Two complementary metrics, Soft IoU-S and Soft IoU-NS, measure reference-generation agreement for facts with and without identified evidence support. On 200 MIMIC-III admissions, 18.8% of physician reference facts have identified support in the prior evidence, compared with 65.4% and 63.5% of facts generated by a general-purpose and a medically specialized model. Overlap is substantially higher for supported content (0.255 and 0.236) than for content without identified support (0.030 and 0.026). The two scores are weakly correlated across admissions (rho = 0.22 and 0.17), and conventional metrics do not consistently capture both dimensions. FairSumEval distinguishes evidence-grounded summarization from agreement on content without identified support, revealing what models omit and add in each category.

preprintarxiv

LLMs and the Homogenization of Teaching

Mei Tan, Hyunji Nam, Hari Subramonyam +1 authors

Mei Tan, Hyunji Nam, Hari Subramonyam, Dorottya Demszky

First announced by arXiv: 2026-11-13

language
Abstract from arxiv

Large language models (LLMs) are known to lack diversity in open-ended text generation, a limitation that is especially consequential in education, where homogeneity can narrow the kinds of thinking and disciplinary work students are invited to do. We study diversity in two educational generation tasks: math items and essay feedback. We compare frontier and open-weight models with human-authored reference distributions using lexical, semantic, and domain-grounded measures. For math, we examine the mathematical activities students are asked to perform; for feedback, we examine the revision work requested and the rhetorical strategies used to communicate it. We evaluate direct prompting, two diversity prompting methods, and post-training. Across both tasks, directly prompted model outputs are markedly less diverse than human-authored content--78% higher on lexical (Self-BLEU) and 51% higher on semantic (embedding) similarity--and represent a narrower range of instructional forms. Diversity prompting and post-training increase variation largely through surface features without broadening the instructional repertoire. Models inherit and amplify the most normative routines of human instruction. Overcoming mode collapse in education will require richer instructional data and investment in teachers' creative capacity and professional judgment.

preprintarxiv

Neural Compression for Scientific Particle Data: Perspectives on How Learning Can Help

Tian Qiu, Carolina Cuesta-Lazaro, Sheng Di +1 authors

Tian Qiu, Carolina Cuesta-Lazaro, Sheng Di, Haewon Jeong

First announced by arXiv: 2026-11-13

ai for sciencevision
Abstract from arxiv

As particle-based simulations grow in scale, efficient compression of scientific data becomes critical. Neural point cloud compression (PCC) methods excel on computer vision (CV) benchmarks, but do they transfer to scientific particle data? We highlight that scientific particles differ from CV point clouds in fractal dimension: CV point clouds are locally planar, but particles are not. This key difference limits the effectiveness of the octree data structure that underpins most PCC methods. Neural codecs achieve only marginal bitrate reductions over the traditional G-PCC baseline on scientific data, revealing a significant performance gap. Supported by an evaluation spanning domains of cosmology, molecular dynamics, and plasma physics, we outline where the structure of scientific data limits current neural approaches and where learning can help in the future.

preprintarxiv

Learning Low-Latency Action Verification for Frozen Vision-Language-Action Models

Wen-Han Hsieh, Ken Jiankun Zheng, Chia-Chien Chen +2 authors

Wen-Han Hsieh, Ken Jiankun Zheng, Chia-Chien Chen, David M. Chan, Trevor Darrell

First announced by arXiv: 2026-11-13

reasoning agentsroboticsvision
Abstract from arxiv

Vision-language-action (VLA) models have made significant progress in general-purpose robotic manipulation. However, sampled action sequences can be ineffective, while evaluating alternatives with computationally expensive verifiers adds latency at each policy query. In this work, we propose HDH, a lightweight framework that uses hindsight from a frozen VLA's own deployments to improve action verification. The framework has three stages. First, we deploy the policy and collect its autonomous rollouts. We then train a compact action verifier to assign each action chunk a predicted task-success score. Finally, the frozen VLA generates action candidates, and the verifier selects the candidate with the highest score for execution. Across four LIBERO-10 tasks, HDH achieves an average success rate of 82.3\% and on three real-world manipulation tasks, HDH improves average success over the baseline policy by 22.6\%.

preprintarxiv

Sophon: A Procedural Diagnostic for Spatial Reasoning in Vision-Language Models

J. Zach Gazak, Ryan Swindle, Justin Fletcher

J. Zach Gazak, Ryan Swindle, Justin Fletcher

First announced by arXiv: 2026-11-13

learning optimizationvision
Abstract from arxiv

Open-weight Vision-Language Models (VLMs) score 85-94% on spatial benchmarks (BLINK Spatial, What's Up CLEVR) yet collapse to chance on simple geometric questions when object, scene, and texture priors are stripped away. We introduce and open-source Sophon, a procedurally generated diagnostic of paired colored point-cloud stimuli across three coordinate systems (1D linear, 2D Cartesian, 2D polar) and four question types (binary, categorical, comparative, quantitative). Across four open architectures (LLaVA-NeXT, LLaVA-OneVision, Qwen2.5-VL, Idefics3), accuracy clusters at 47-52% while frontier models (Opus 4.6, Gemini 2.5 Pro, GPT-5.4) clear 77-83% and humans reach 84-94%. We use linear probes and low-rank adaptation (LoRA) to trace the open-weight failures. Probes of the vision encoder (ViT) and language decoder (LLM) recover Sophon spatial information at 83-85% and 80-86% accuracy, respectively, and the language tower's reasoning circuit answers text-only controls (point clouds as sets of coordinates) at +40 pp over visual stimuli. Language tower per-layer attention attribution localizes visual integration to mid-stack layers. At readout, models elevate the correct answer-set tokens but select among them randomly. We localize the mechanism by tuning individual and grouped components across every architecture's stack, including the vision tower (ViT workspace and reader), the projector, and the language tower (LLM workspace and reader). These experiments demonstrate that this mid-stack workspace holds circuitry capable of organizing the spatial signal so that the frozen reader can decode it. LoRA fine-tuning adjusts that circuitry such that 7B-class VLMs answer Sophon spatial reasoning questions at frontier aggregate accuracy. This work suggests that these models do not have a capability gap for spatial reasoning but suffer from a routing bottleneck that targeted fine-tuning repairs.

preprintarxiv

SPIRE: Task-agnostic embodied continual learning through autonomous situation inference and causal motor-primitive reuse

Yiyang Jia, Mingyang Lyu, Gongce Cao +3 authors

Yiyang Jia, Mingyang Lyu, Gongce Cao, Feifei Zhao, Yinqian Sun, Yi Zeng

First announced by arXiv: 2026-11-13

learning optimizationrobotics
Abstract from arxiv

Embodied continual learning requires controllers to infer changing situations, retain prior behavior and reuse motor skills without task labels or known boundaries. The challenge is deciding when to recover earlier behavior and which skills can support new learning. We introduce Situation-guided Primitive Integration for Recurrent Embodiment (SPIRE), a recurrent-spiking system that infers latent contexts from action-outcome evidence and learns context-specific motor actors online. Mature actors become protected sources that combine with a target-context residual to drive recurrent-spiking motor populations. Across coordinate, three-joint and hidden-force control, SPIRE maintained control, autonomously inferred situations, retained prior behavior and accelerated acquisition of an unseen source composition, reducing mean interactions to criterion by approximately 40% relative to matched fresh controllers. Rewinding compatible source actors impaired early acquisition in every matched intervention, demonstrating causal reuse. Against seven recurrent continual-learning and modular-control comparators, SPIRE achieved 2.4 times the mean prior-learning advantage of the strongest eligible task-free alternative selected for each paired stream. These results demonstrate autonomous recovery and causal reuse of motor skills for new learning in continuous embodied control.

preprintarxiv

AutoAssumption: Searching over Assumptions Discovers ML-Systems Algorithms

Igor Sadalski, Minghao Guo, Benjamin Jones +1 authors

Igor Sadalski, Minghao Guo, Benjamin Jones, Wojciech Matusik

First announced by arXiv: 2026-11-13

learning optimizationreasoning agents
Abstract from arxiv

Many advances in ML systems came from breaking an assumption that earlier methods took for granted: speculative decoding dropped the assumption that a model needs one forward pass per token. Yet algorithms rarely state their assumptions, so experts and discovery harnesses search over programs, where one edit changes several design choices and gains are difficult to attribute. Searching over the assumptions themselves remains largely unexplored. Writing algorithms as proofs makes their assumptions explicit and turns algorithm design into a search over a small, structured space that an LLM can explore and a proof checker can audit. We formalize 89 algorithms from KV-cache eviction, weight pruning and speculative decoding as proof graphs with explicit assumptions, checked in Lean. Our harness, AutoAssumption, relaxes one assumption of a published method, re-derives and re-proves it, and generates its code. In KV-cache eviction and weight pruning, this search improved on every published algorithm we ran and produced state-of-the-art algorithms. The eviction algorithm has the highest RULER-4K score of the published methods we ran, and the pruning algorithm the lowest perplexity of thirteen on four LLaMA models. At equal budget, the harness scores within a point of OpenEvolve, an open-source implementation of AlphaEvolve, with a lower standard deviation across repeats. Future work will formalize more fields and combine them through the principal assumptions they share.

preprintarxiv

Open Challenges and Opportunities in Human-Agent Systems

Reuth Mirsky, Sarath Sreedharan, Nitay Alon +7 authors

Reuth Mirsky, Sarath Sreedharan, Nitay Alon, Stefan Sarkadi, Munindar P. Singh, Liz Sonenberg, Peter Stone, Silvia Tulli, Stylianos Loukas Vasileiou, Shlomo Zilberstein

First announced by arXiv: 2026-11-13

reasoning agents
Abstract from arxiv

Agentic AI systems increasingly undertake extended tasks and act on people's behalf, making human-agent interaction central to their effectiveness. Successful interaction requires understanding human perspectives, coordinating complementary capabilities, and maintaining productive relationships within socio-technical environments. This paper examines how foundations developed by the Autonomous Agents and Multi-Agent Systems (AAMAS) community inform these demands. We consider interactions ranging from individual human-agent partnerships to teams and larger groups, where participants differ in their roles, decision authority, capabilities, and objectives. We organize computational approaches around the connected challenges of understanding others and deciding how to interact with them. Theory of Mind, belief-desire-intention architectures, goal and plan recognition, inverse planning, and recursive reasoning provide mechanisms for representing and inferring partially observable mental states. Observer-aware planning, model reconciliation, and decision-theoretic communication connect these inferences to action, explanation, and information exchange. Together with shared autonomy, argumentation, and cooperative inverse reinforcement learning, these approaches offer foundations for agentic systems that account for evolving human knowledge, intentions, and expectations. We argue that progress requires treating human interaction as an ongoing decision problem shaped by uncertainty, interdependence, and potentially conflicting objectives.

preprintarxiv

Generalize Then Personalize: Wearable-Based Heart Rate Forecasting Under Scarce Ruck-March Data

Muntaqim Ahmed Raju, Ruizhe Ma, John W. Ramsay +1 authors

Muntaqim Ahmed Raju, Ruizhe Ma, John W. Ramsay, David J. Cornell

First announced by arXiv: 2026-11-13

learning optimization
Abstract from arxiv

Heart rate (HR) forecasting may provide advance information about near-future cardiovascular response, potentially supporting physiological monitoring and workload-management decisions during ruck marching. However, robust model development is constrained by the difficulty of collecting large-scale ruck-march datasets, creating a scarce-data problem in which models must generalize across participants while adapting to individuals from limited personal data. To address this problem, we benchmark population-only training (POT), personalized-from-scratch training (PFST), and population-pretrained fine-tuning (PPFT) for HR forecasting during loaded ruck marching using data from eight active-duty U.S. Army medics across thirteen sessions. LSTM, Attention-LSTM, and CNN-LSTM backbones were evaluated against a persistence baseline at 30 s and 60 s horizons. Across all neural architectures and horizons, PPFT yielded lower mean MAE, RMSE, and endpoint MAE than both POT and PFST. Participant-level paired analysis after averaging across three training seeds showed that, relative to PFST, PPFT reduced trajectory MAE by 2.22-4.30 bpm and endpoint MAE by 1.24-2.88 bpm. All six trajectory-MAE comparisons remained significant after Holm adjustment ($p_{adj}$ = 0.0469), while four of six endpoint-MAE comparisons remained significant after adjustment. At 30 s, persistence achieved the lowest MAE (3.555 $\pm$ 1.274 bpm), while CNN-LSTM PPFT achieved the lowest RMSE (4.669 $\pm$ 1.916 bpm); LSTM PPFT achieved the lowest endpoint MAE (5.093 $\pm$ 2.365 bpm). At 60 s, Attention-LSTM PPFT achieved the lowest MAE (5.573 $\pm$ 1.939 bpm) and RMSE (7.134 $\pm$ 2.260 bpm), while CNN-LSTM PPFT achieved the lowest endpoint MAE (7.553 $\pm$ 2.357 bpm). These results suggest that population pretraining followed by participant-specific fine-tuning may provide a practical strategy for scarce-data wearable HR forecasting.

preprintarxiv

Effective Teacher Support for Reasoning Distillation

Lavinia Ghita, Dhruv Desai

Lavinia Ghita, Dhruv Desai

First announced by arXiv: 2026-11-13

languagelearning optimizationreasoning agents
Abstract from arxiv

Forward-KL language-model distillation normally computes dense teacher and student scores over the full vocabulary at every supervised position, making training and stored teacher targets expensive. We instead distill from the teacher's highest-probability vocabulary fraction, renormalizing the retained mass as we sweep support from 100% to 0.01%. Our main study distills Qwen3 4B -> 1.7B. We confirm the broader support-reduction finding on larger Qwen3 8B -> 4B and Nemotron Nano 12B -> 9B pairs and under answer-only supervision. We evaluate generation primarily on MMLU-Pro and secondarily on GPQA-Diamond. Across the Qwen3 settings, retaining 0.01% of the vocabulary causes no observed MMLU-Pro degradation relative to full-vocabulary distillation, although disagreement with the complete teacher distribution increases. Nemotron Nano matches full-vocabulary MMLU-Pro at 0.1%, but smaller fractions fail to emit the closing reasoning tag, showing that the reliable cutoff depends on the model family. At 0.01% support, caching the retained next-token scores from the teacher forward pass cuts training step time by 38%, while restricting student output computation to those tokens cuts peak memory by 24%; neither causes observed MMLU-Pro degradation.