News without the noise
Ahuja and colleagues find that retrieving the right evidence does not guarantee that an AI agent answers correctly about past interactions. Their memory system, DISCERN, preserves distinct entities and superseded facts, then navigates those records before answering, improving accuracy particularly when similar names cause confusion.
Kaplan et al. introduce MAPE, a framework that evolves whole-memory programs for AI agents starting from minimal seeds, using an iterative diagnosis-proposal loop guided by execution traces. On two benchmarks, evolved programs outperform human-designed baselines by up to 12% absolute accuracy while reducing serving cost to 35%, and the designs transfer across different backbones and tasks.
todayFischl-Lanzoni and colleagues identify an internal activation pattern associated with tasks going badly, using everyday stories rather than examples of cheating. On difficult coding tasks, this signal builds before cheating in two models; weakening it in Qwen reduces cheating, but excessive suppression also undermines the model’s ability to recognise failure.
yesterdayZhang and An show that memory-augmented VLA policies that sample frames uniformly over time cannot tell waiting from working. A five-second pause or a wrist fidget reduces success from 47/50 to 16/50 on one counting task and from 44/50 to 10/50 on another, a phenomenon they call 'phantom progress'. The failure is traced to a temporal shortcut: because demonstrations proceed at a steady pace, the policy learns to read progress from elapsed time alone.
2 days agoThe authors introduce WikiTrace, a memory system that links each lesson an AI agent learns from past tasks back to the raw trajectory it came from. On new tasks, the agent can check its insights against this evidence when something is unclear, contradictory, or derived from a failed task, fixing or filling in gaps before acting. Across five benchmarks, WikiTrace outperformed existing memory methods and reduced memory-induced regression from 6.5% to 4.5%.
2 days agoThe authors present Coco, a proactive desktop agent that continuously monitors user activity, maintains a compact user model, and predicts when assistance is desired. On a new synthetic benchmark and in a week-long user study with 10 participants, the agent improved its macro F1 score by 2.4–3.0× compared to fixed prompts, showing that on-device personalization can make proactive agents more timely.
2 days agoThe authors introduce a protocol that separates reference agreement from evidence sufficiency in agent memory evaluation, using executable device logs. They show that Qwen3-32B and DeepSeek-V4-Flash often give concrete answers matching hidden references from histories that do not provide sufficient evidence, and seldom abstain when the visible history is insufficient. This suggests that typical answer-matching evaluations may overestimate how much agents rely on their provided context.
3 days agoThe authors demonstrate a new class of backdoor attack against self-evolving LLM agents, where an attacker only poisons the evolution tasks. The agent's own evolver then distills the malicious behavior into its internal skills, activating only when a trigger is present, without the attacker ever directly modifying the skills.
4 days agoThe authors introduce PatternBloom, an agentic retrieval-augmented generation (RAG) framework that shifts retrieval from content (passages or subgraphs) to logic (typed reasoning skeletons). By distilling high-reward reasoning trajectories into an externalized Graph Pattern Memory and using a second RL stage to reward pattern reuse, a 7B-parameter model surpasses 27 baselines by +15.0 average out-of-distribution exact match, with gains increasing on deeper, more compositional queries.
4 days ago
