Methods worth understanding

AI Techniques Radar

New AI methods, the evidence behind them, and how to put them to work.

Latest weekly report

AI Techniques Weekly

Methods worth understanding, their results, and practical limitations.

Read the report
Reset

Latest discoveries · 115

Last refreshed 2026-11-10 · Some sources need attention

Training & fine-tuning

Use a lightweight fine-tuning library that overwrites parts of the model’s training code and replaces PyTorch modules with custom Triton kernels derived from backpropagation steps, to speed up QLoRA-style fine-tuning while remaining compatible with the Hugging Face training stack.

AI summary

Inference efficiency

Use a small “draft” model to propose multiple candidate tokens, then have the larger “target” model verify all candidates in parallel via a single forward pass. In the reported approach, both the draft and target are quantized (with choices compared between 8-bit SmoothQuant and 4-bit weight-only quantization) to improve the balance between memory bandwidth and compute overhead on Intel Xeon.

AI summary

Agents & tool use

Back to Articles Evaluating Audio Reasoning with Big Bench Audio Published December 20, 2024 Update on GitHub Upvote

Source excerpt

Retrieval & memory

A retrieval method that reframes a multi-turn question as a set of evidence requirements, retrieves candidate dialogue turns, checks whether the retrieved items jointly cover the required evidence (using source identifiers), and then selects turns as a coordinated set to support later answer generation.

AI summary

Inference efficiency

Automatically refines inference precision at runtime by selecting per-layer, per-step formats for weights, activations, and the KV cache. A lightweight policy chooses the precision using runtime signals such as next-token distribution entropy and the magnitude of each layer’s hidden-state updates. Precision formats are generated from a shared floating-point base via multi-scale exponent/mantissa refinement, enabling multiple reduced-precision variants from a single underlying format.

AI summary

Training & fine-tuning

Improve a motion tokenizer used with diffusion by adding a spectral-focused adaptation step during text-conditioned training. A pretrained diffusion model acts as a critic to suppress undesired high-frequency energy in the tokenizer’s latent space, aiming to reduce a generation–reconstruction mismatch attributed to spectral bias and temporal artifacts.

AI summary

Training & fine-tuning

Build a benchmark that scores multiple candidate research reports using query-specific, golden rubrics, then evaluate reward models (and LLM judges) on ranking-style preference data (e.g., pairs/triplets/quadruplets) rather than only direct scalar scoring, with dataset quality controlled via multi-round filtering and human verification.

AI summary

Reasoning & planning

Define each patient (or group) by the feature space they come from, then train with an objective that focuses on the worst-performing group in terms of excess loss over that group's estimated best achievable loss. Use group-specific encoders that map each group's features into a shared latent space, where learnable Gaussian anchors cluster patients by outcome to enable knowledge transfer even when feature spaces are disjoint.

AI summary

Inference efficiency

A test-time inference framework that performs iterative spatial reasoning by selecting spatial transformations, executing them in a latent world model, and feeding aligned latent intermediate evidence back into the VLM—avoiding repeated latent-to-pixel-to-latent decoding. It further uses process-verifiable reinforcement learning to reward actions only when both the action and the resulting latent transition are valid.

AI summary

Training & fine-tuning

Learn cross-embodiment correspondences by training self-supervised temporal representations with a cycle-consistency objective, then use the resulting alignments to co-train an autoregressive video-action diffusion model with behavior-aligned objectives and an additional human-to-robot style-transfer objective.

AI summary