Persistent skills, harness evolution, and calibrated agent action

Liyan Tang, Cyrus Rashtchian, Chun-Sung Ferng, Andrew Tomkins, Da-Cheng Juan, Tu Vu

WikiSkill co-evolves executable agent skills with a persistent wiki that consolidates raw execution traces into reusable knowledge. Across benchmarks it outperforms prior skill-evolution methods; larger models gain more from evolved skills, smaller models with skills can beat larger models without them, and skills transfer across model families, with the wiki ablation confirming that persistent accumulation is essential.

Yufan Wu, Yinghui He, Zhengyi Hu, Lang Wei, Ruichen Li, Qifan Yang, Ting Zhu

CritICL treats structured failure modes of weaker same-family models as critique-based in-context guidance for stronger models at inference time. Dynamic and static variants outperform standard ICL and match or beat test-time scaling while using far fewer generations and tokens, showing that cross-scale failure patterns are a cheap source of weak-to-strong generalization.

Aozhe Wang, Zhengxi Lu, Jianze Wang, Shangke Lv, Ying Liu, Weiming Lu, Jun Xiao, Yueting Zhuang, Hua Yang, Qianglong Chen, Yongliang Shen

TTPO enables label-free test-time training by an asymmetric objective: OPSD distills rollouts that agree with majority-vote pseudo-labels while Grouped RL penalizes disagreeing ones, with token-level selection refining both. Without ground truth it matches supervised OPSD on competition math, lifts Qwen3-1.7B substantially in TTT, and generalizes across tasks, remaining robust even when pseudo-labels are frequently wrong.

Xingshan Zeng, Zishan Xu, Boju Zhang, Yuzhou Wu, Lingzhi Wang, Jianghao Lin, Liangyou Li, Yasheng Wang, Lifeng Shang, Xin Jiang, Weinan Zhang, Yong Yu, Qun Liu, Weiwen Liu

The paper factorizes agentic data as (environment, task, trajectory, verifier) and frames generation as constrained distribution design under an Accuracy-Complexity-divErsity (ACE) lens. Accuracy enforces grounded consistency, complexity places mass relative to a declared learner, and diversity controls coverage; the survey shows the field shifting toward execution-grounded, learner-relative, non-redundant experience rather than raw volume.

Jinghan Xu, Yikai Zhang, Aili Chen, Weiyuan Li, Jiaqing Liang, Deqing Yang

HarnessLens jointly explores tasks and harness components, proposes modifications from trajectories, and verifies only on behavior-relevant tasks under an attributable-evidence gate. On three harnesses and four benchmarks it raises held-out performance 7.6-13.6% while using far less evaluation budget than full propose-and-verify baselines, showing behavior-aware selective verification enables reliable harness evolution under tight interaction limits.

Pranav Aggarwal

Across 12 frontier models, professionally packaged evidence (even fully fabricated numbers) sharply raises commitment to directional answers on provably unknowable questions, while stated probabilities and knowability classifications remain largely unchanged. The failure is localized to the act/don't-act gate; SFT on a few hundred synthetic irreducible cases drives commitment to zero and transfers, yet the gate is fragile under rigid response formats that remove reasoning room.

Yunpeng Ba, Zhi Zheng, Yue Xie, Jiaqing Li, Xialiang Tong, Tao Zhong, Mingxuan Yuan, Zhichao Lu, Xuyang Wu, Zhenkun Wang

Analysis shows Evolution Strategies produce higher verifier-projected diversity than GRPO, yielding better Pass@K while still improving Pass@1 and avoiding entropy collapse. Gains concentrate in a sparse subset of large-magnitude updates despite large parameter drift, without catastrophic forgetting; a sequential GRPO-then-ES recipe combines strengths, positioning ES as a distinct reasoning post-training paradigm rather than a mere memory-efficient substitute.

Orion Reblitz-Richardson

Linear probes for the six Moral Foundations Theory categories span near-maximal independent dimensions yet share a positive moral-specific common component that appears early in pre-training. Dilemma directions partially compose from their foundation components while mostly encoding conflict-specific structure, so the model represents moral tension rather than a pre-resolved judgment and does not recover MFT's individualizing/binding split.

Jackie Baek

With a single untuned prompt and a Python sandbox, frontier LLMs match or beat specialized methods on inventory control, queuing network control, and assortment optimization, both for single instances and when returning a fixed algorithm for a problem class before seeing test instances. Capability improved sharply across models released months apart, establishing untuned LLM algorithm design as a serious empirical baseline for well-specified OR problems.

Shengzhuang Chen, Jerrod Parker, Yejin Bang, Andrew M. Bean, Nabeel Seedat, Stefan Winzeck, Daniil Glazko, Jannik Zgraggen, Fangyi Yu, Scott Arnott, Dietrich Trautmann, Luca Ciuffreda, Guglielmo Bonifazi, Davide Romano, Bradley Bell, Kirsty Fielding, Daniele Giofrè, Tom Zielund, Ipshita Chatterjee, Sneha Murthy Ghantasala, Manpreet Nanreh, John Scoville, Maciej Sakowicz, Wassim Seifeddine, Lukas Thede, Jonathan Richard Schwarz

Thomson shows that careful continual mid- and post-training of open-weight models, with plasticity/stability safeguards and minimal high-impact interventions, can produce frontier-competitive general-purpose performance on agentic, safety, legal, tax, multilingual, and deep-research tasks at far lower cost than training from scratch. Results exhibit a π-shaped pattern of broad gains including untargeted capabilities while nearly eliminating catastrophic forgetting, supporting practical SovereignAI ownership of model, tools, values, and data.