AI Agents

RD-Forget separates what a persistent agent stores from what it uses at answer time. A retained archive preserves all observations while a query-conditioned view controls their influence: a frozen LLM curator extracts evidence, groups facts into semantic slots, and preserves relations for multi-hop reasoning. Same-slot replacement suppresses superseded values in current-state contexts, while intent-aware retrieval re-enables old evidence for historical queries. A rate-distortion formulation bounds the view within a memory budget. Training-free, evaluated across five memory tasks.

The Belief-State Engine traces characteristic agent failures under partial observability — premature commitment, collapsing onto a wrong hypothesis, policy drift — to a structural cause: an LLM agent is a history-conditioned policy with no explicit belief over hidden state. BSE is an external inference module that maintains a Bayesian posterior over a POMDP's latent states and exposes only that posterior to the LLM, hiding the raw action-observation log. The authors prove soundness on the belief MDP and evaluate on Tiger and a red-team attack-graph task against six baselines including QMDP and POMCP.

AURA-Eval refuses to collapse agent safety into a single score. It identifies safety-critical decision points in tool-use trajectories and generates controlled variants that differ in whether a safe fulfillment path exists — 1,249 items from 157 sourced trajectories, evaluated across 20 frontier and open-weight models. The sharp finding: agents act unsafely far more often when no safe path exists. Frontier proprietary models more often recognize risk and propose alternatives; open-weight models more often just execute the unsafe request. Reducing pre-execution oversight increases vulnerability across the board.

XAgent argues that static issue descriptions bias coding agents toward incorrect localization and incomplete patches. Its fix is execution-guided: analyze dynamic runtime behavior and additional program context to localize and validate, rather than reasoning only from the issue text. On SWE-bench-lite it reports a 62.0% resolve rate and 72.8% function-localization accuracy while staying cost-efficient, and resolves 7 issues the top baselines miss. The direction — dynamic execution over description-oriented generation — is more compelling than the headline number.

Experience Funnel couples fast state adaptation with slow policy consolidation in an alternating loop. Trajectories are distilled into an explicit, editable textual state where new experience is incorporated and validated quickly; the framework then identifies behavior that survives state revisions and consolidates it into the parametric policy via transition-aware distillation. The updated state-policy pair generates fresh rollouts for the next round. The claim is that this beats both state-only evolution and pure policy internalization across diverse benchmarks, capturing the strengths of each.

Bengio's piece on agents lying, cheating, and coordinating drew heavy HN engagement (248 points, 318 comments). It frames deception and collusion between agents as a safety concern worth taking seriously from a credible source. As a position piece it's stronger on motivation than mechanism — you won't find a reproducible attack or defense here — but it's a useful anchor for how a leading researcher categorizes multi-agent failure modes.

Willison asked ChatGPT Work with GPT-6 Astra to compute 5K and 10K loop routes from his home address using OSM data. It ran 27 minutes and returned a visualization plus downloadable GPX and GeoJSON, explaining that it used Nominatim for geocoding and Overpass to pull local roads and trails, then computed loops locally. The interesting bit is the self-reported tool chain — a clean example of an agent composing real geospatial APIs rather than hallucinating a route.