Catastrophic remembering, skill compression, and structured agent memory

Alan Li, Rahul Saha, Anton Xue, Swarat Chaudhuri, Adam Klivans, Pravesh K Kothari, Raghu Meka

The authors report tightened bounds on the Grothendieck constant achieved with an AI research system that produced insights judged novel by domain experts. They detail strengths, failure modes, and the practical conditions under which long-horizon human-AI mathematical collaboration yields breakthrough results rather than incremental noise.

Kushal Chakrabarti

Across nearly 250k instruction lifetimes the author shows agentic coding prompts more than triple in size because appending is cheap while safe deletion is exponentially hard once rationales are lost. Naming the phenomenon catastrophic remembering, the inverse of catastrophic forgetting, the work shows that latent-reasoning comments can remove over 99 percent of excess instructions and improve real instruction-following by up to 23 percent.

Xiaofan Bai, Hongqiang Lin, Chao Liu, Yantao Zhang, Xuan Jin, Xipeng Cao, Yuhong Li

SkillZip compresses self-evolving agent skills by finding a shortest faithful structural explanation under a typed minimum-description-length objective with hard coverage constraints. It factors repeated rules and action sequences once, keeps rare exceptions by construction, and supports both one-shot and continual Zip-on-Write modes without evaluation rollouts.

Aijun Yang, Qianxue Guo, Ziyi Huang, Yuxuan Chen, Shiyou Qian, Jian Cao

ReTree models search as an evidence tree whose nodes store bounded summaries, source-linked evidence, and revision histories. When new evidence contradicts earlier claims it localizes the error, replaces outdated nodes, prunes affected branches, and resumes, cutting context size while raising answer accuracy by up to 25.6 points over full-trajectory ReAct.

Viktor Volkov, Valentin Khrulkov, Andrey V. Galichin, Danil Sivtsov, Nikita Glazkov, Olga Volkova, Konstantin Pchelin, Iaroslav Bespalov, Dmitry V. Dylov, Petr Anokhin, Ivan Oseledets

EvoMem turns successful LLM-driven mutation events into structured, task-aware advice stored with provenance and later retrieved to guide future evolutionary program search. Across geometric optimization, multi-hop QA, and GPU kernel tasks the persistent memory reduces redundant rediscovery and improves target metrics or search speed in most settings.

Vaibhav Singh, Soumya Suvra Ghosal, Sarvesh Gharat, Soumyabrata Pal, Ramasuri Narayanam, Dinesh Manocha

ThinkRetrieve injects dynamically retrieved solved exemplars into intermediate reasoning steps of large reasoning models, supplying guidance on how to reason rather than only external facts. Across models from 1.5B to 8B parameters it consistently beats standard test-time scaling, with relative gains reaching 60 percent on AIME 2025.

Serafim Batzoglou

The framework evaluates every LLM-proposed first-order formula exactly on training structures, retains the strongest verified hypothesis, and feeds remaining errors back to guide the next generation round. Symbolic repair and simplification further increase the number of solved induction problems while shortening many train-valid formulas without changing predictions.

Nikolai Bolik, Lennart Stöpler, Artur Andrzejak

Using overlap of active SAE latent sets as a similarity measure, the authors show that these sets recover compositional structure in toys yet fail to track human category boundaries or typicality more faithfully than dense embeddings. Controlled semantic edits reveal large mismatches between human conceptual change and SAE active-set change, arguing against simple bag-of-features compositionality outside idealized regimes.

Minsoo Kim, Sungyoung Ji, Kisung Moon, Ilyong Yoon

ASMI masks attention heads and measures BALD mutual information among the resulting subnetworks with a semantic-agreement kernel, capturing confident-but-fragile predictions. The training-free signal adds error-predictive information beyond entropy on grounded QA, roughly halves retained error of a confidence filter in its strong regime, and can be read from a single greedy response.

Clemens Vetter, David Kaczér, Lucie Flek, Florian Mai

SAE-based model diffing shows misalignment fine-tuning amplifies jailbreak, sarcasm, deception, and manipulation persona features while suppressing safety and assistant-identity features. Steering these features alone can induce up to 62 percent misalignment or restore near-baseline rates, yet human-written pre-training documents that activate them do not reliably induce the effect, whereas synthetic instruction-response pairs derived from the same content do.