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

Reasoning & planning

Use a regime-aware benchmark/audit protocol that evaluates LLM judges on multiple, explicitly separated response regimes—e.g., verifiable constraint-violation evidence, preference judgments with “four-of-five” denominators, and within-panel disagreement—rather than collapsing everything into a single score. Add tie-band calibration and measure call-to-call variation so judge reliability is assessed across rubric regimes.

AI summary

Training & fine-tuning

Use a differentiable MPC layer that blends SQP with a log-barrier-smoothed ADMM inner solver to obtain smoother gradients and exact sensitivities when the inner solve converges. Then train the resulting MPC policy with a first-order actor-critic method (Short-Horizon Actor-Critic) that reuses rollout MPC solutions to avoid repeatedly solving/evaluating MPC during training.

AI summary

Inference efficiency

Maintain client-specific “Age of Intelligence” that measures how much a cached model’s outputs are expected to distort action utility relative to a newer version. Synchronize a cached model only when the expected stale-decision regret outweighs the refresh cost, and estimate output drift by accumulating Kullback–Leibler–divergence changes from client-specific unlabeled probe data rather than requiring labels.

AI summary

Reasoning & planning

Use a sandbox benchmark that supplies configurable structural equation models with explicit resource constraints, then evaluate an agent’s ability to plan experiments, iterate through sequential observations, analyze results, and decide when to stop and conclude—without providing domain-specific semantic cues.

AI summary

Reasoning & planning

Use distributional RL with quantile-based value modeling (e.g., IQN) and replace fixed/random scalar CVaR training with preference augmentation that shifts quantile-weight distributions. Train via preference-guided exploration over diverse risk preferences, then compute a conservative policy at deployment by optimizing against a worst-case preference distribution rather than a hand-tuned single risk scalar.

AI summary

Reasoning & planning

The technique studies whether a multimodal attention imbalance arises because one modality’s tokens enter the model with lower semantic density. It then uses a token-position replacement intervention—swapping vision token positions with semantically closest text-token counterparts—to probe whether early-layer attention shifts toward the replacement positions, indicating that the model is initially “processing text” before effectively combining with vision.

AI summary