Inference efficiency
Evaluate conversational voice agents by running complete multi-turn spoken interactions in a bot-to-bot audio simulation, then score each dialogue along two coupled dimensions: task accuracy and spoken conversational experience. Use deterministic, tool-log-based checks for objective completion, plus LLM/LALM judge metrics for faithfulness and audio-level speech fidelity, and experience metrics for conciseness, progression, and turn-taking timing.
AI summary
Inference efficiency
Run an agent in an isolated, stateful sandbox where the evaluation judges the terminal backend state and extracted side effects (e.g., what records were changed) against an executable required end state, rather than trusting the agent’s final text or tool-call formatting. Repeat each task multiple times from a clean initial backend to measure consistency.
AI summary
Reasoning & planning
Split an ultra-high-resolution remote-sensing visual QA system into (1) a region/exploration policy that uses a spatial pointer verifier to decide whether search is needed, (2) evidence-aware view scoring that runs lightweight decision heads over compact candidate views using frozen visual representations (without generating textual region judgments), and (3) format-specific confidence calibration to route between candidate answers generated from different evidence presentation formats (e.g., memory vs overview).
AI summary
Inference efficiency
A parameter-efficient method that adapts a frozen CLIP video backbone by specializing its features along time. It uses (1) learnable temporal positional embeddings to align temporal phase and (2) a local motion transition operator implemented with depth-wise convolution to model short motion changes between nearby frames.
AI summary
Inference efficiency
Evaluate SSL methods using both ranking metrics and reliability metrics to expose unreliable confidence under domain shift, with repeated runs and logged diagnostics across label regimes and datasets.
AI summary
Inference efficiency
Introduce a quality-governed encoder-decoder segmentation model that estimates slice quality from contrast and signal-to-noise (SNR) statistics, then uses learnable quality gates to fine-tune encoder features before decoding. This aims to reduce weak-contrast or noisy skip-connection effects when tumor boundaries are heterogeneous.
AI summary
Reasoning & planning
Model anomaly persistence during diffusion-based reconstruction as a deterministic recursion that yields a continuous “survival law.” The law describes how the anomaly decays as noise decreases, controlled by a retention parameter tied to which internal states the sampler visits. This framing treats leak growth with steps as a form of exposure bias and relates trade-offs between leak and reconstruction/detail error.
AI summary
Reasoning & planning
A framework for generating reports that combines (1) dense memory summaries and (2) visually gated retrieved evidence, then performs case-specific fusion across multiple evidence views. It uses a prompt design that separates finding mentions into positive, negative, and uncertain statuses, to better align generated text with clinically consistent cues.
AI summary
Training & fine-tuning
Train a multi-block vision transformer so that intermediate blocks produce discriminative identity embeddings suitable for similarity-based matching. Each training iteration selects a random intermediate block and applies the same embedding head and classification loss used for the final block, using a weighted margin-penalty softmax auxiliary objective alongside the main final-layer loss.
AI summary
Inference efficiency
QCVS is an inference-time intervention method for multimodal (vision-language) reasoning that aims to separate query-relevant reasoning effects from low-level visual noise. It uses a question-contrastive, double-differential activation probing: it compares activation changes caused by spatial/layout perturbations under the target query against changes from a generic (task-agnostic) description reference, so that visual artifacts not specific to the question cancel out. A dual-threshold gating mechanism then decides whether to apply a bounded hidden-state update, based on relative interaction strength and signal amplitude, to avoid unnecessary edits.
AI summary
Inference efficiency
Use a preconditioned implicit-midpoint discretization of Langevin dynamics, where an implicit midpoint update (via a Cayley-transform identity) is combined with matrix preconditioning to avoid the stationary-variance inflation attributed to preconditioned Euler-Maruyama. Optionally, replace the non-smooth regularization with a Moreau-envelope-based proximal form to handle non-smooth objectives.
AI summary
Retrieval & memory
Use a two-part KV cache system for multi-round, long video generation: (1) compress KV data within each chunk using a channel/video-aware low-rank decomposition whose rank is predicted online to satisfy an energy-retention target; and (2) manage compressed chunks across time with an inter-chunk cache manager that shares basis directions between consecutive chunks and tracks chunk states in a tree so branches can reuse history and later restore for resume/rewind/fork.
AI summary