papers · 2026-05-22

SkillOpt’s sharp move is treating an agent skill as trainable external state, not dressing up reflection loops again. A separate optimizer model converts scored rollouts into bounded add/delete/replace edits, then accepts only edits that improve held-out validation. Deployment adds zero extra inference calls. That is much closer to an optimizer than GEPA, TextGrad, or EvoSkill-style text search. The numbers are hard to ignore: six benchmarks, seven target models, three harnesses, and best-or-tied results across all 52 evaluated cells. On GPT-5.5, it reports +23.5 points in direct chat, +24.8 inside Codex, and +19.1 inside Claude Code. My first checks would be validation/test separation and whether the human, Trace2Skill, GEPA, and EvoSkill baselines were implemented strongly. If those survive, this is a credible path to agent gains without touching weights.

The useful part is not the Shannon metaphor; it is the attempt to fit overtraining and quantization damage with one curve. The paper fits on Pythia up to 6.9B parameters and 180B tokens, then predicts an unseen 12B run up to 307B tokens with pooled R²=0.847. That is a better story for late-training failure than monotonic power laws. I still do not buy “capacity law” yet. A 12B extrapolation sits inside the small-model regime, and Pythia / OLMo2 are not frontier training stacks. Chinchilla mattered because its compute-data tradeoff held across budget scales; this paper is stronger as a perturbation model. Before anyone uses it to plan 70B, MoE, or long-context continued pretraining, it needs cross-architecture evidence.

Agent skill libraries fail at the consumption step, not only at extraction. The paper spans five agentic task domains and lands on an ugly result: model-generated skills help on average, but negative transfer is non-trivial, and utility is independent of model scale or baseline strength. That cuts against the optimistic Voyager / Reflexion line: compress experience into reusable procedures, then agents get better. Here, a strong extractor can be a weak consumer, and the reverse also holds. The proposed meta-skill improves skill quality across domains and reduces negative transfer, but the snippet gives no benchmark numbers. Without those, I don’t buy this as a general agent-memory recipe yet; I buy it as a useful demolition of the naive skill-library story.

All 3 sources use the same title and sit on the arXiv / HF paper-distribution chain; this is synchronized indexing of one v1 paper, not independent market validation. CHRONOS has a clean technical hook: 500 sellers, Recall@10 of 0.937, 161 ms latency, 2.74 QPS, and ε=4.25 at δ=1e-6, tying temporal KG routing, Shapley valuation, and DP budget control into one architecture. I buy the problem framing, but I don’t buy the “data marketplace” strength yet. The paper itself says released valuations remain noise-dominated at that privacy level, and utility mainly comes from public index routing plus adaptive scheduling. For agentic data infra builders, this reads more like a privacy-constrained routing and scheduling prototype than a settlement-ready Shapley marketplace.