Ctx2Skill: Evolving Reusable Skills from Context via Multi-Agent Self-Play
Authors: Shuzheng Si, Haozhe Zhao, Yu Lei, Qingyi Wang, Dingwei Chen, Zhitong Wang, Zhenhailong Wang, Kangyang Luo, Zheng Wang, Gang Chen, Fanchao Qi, Minjia Zhang, Maosong Sun
Institutions: Tsinghua University, DeepLang AI, UIUC, Fudan University, CUHK
Paper: arXiv:2604.27660 (v1 2026-04-30, v2 2026-05-03)
Code: https://github.com/S1s-Z/Ctx2Skill
Keywords: in-context learning, skill extraction, multi-agent self-play, adversarial collapse, cross-time replay
Key points
- Problem: LLMs struggle at genuine *context learning* — inducing new knowledge from a novel context (e.g., an unfamiliar technical manual) to solve tasks. Existing approaches suffer from expensive human annotation and no feedback signal, since in-context tasks lack ground truth to verify whether extracted skills are complete or faithful.
- Approach: Instead of human annotation, Ctx2Skill runs a multi-agent self-play loop where models challenge, solve, judge, and upgrade skills against each other.
- Hard probe: per iteration, the failed task with the fewest rubric passes
- Easy probe: per iteration, the successful task with the fewest rubric passes
- w/o Challenger evolution: 13.8% (-2.7%, largest drop — adversarial pressure is core)
- w/o Cross-Time Replay: 14.7% (-1.8%)
- w/o Hard Probe: 15.2%; w/o Easy Probe: 15.7%; w/o Laplace smoothing: 15.5%
- Merging Proposer+Generator: 15.9%
- GPT-5.1 skills → GPT-4.1: 16.1% (nearly matching GPT-4.1's self-produced 16.5%)
- GPT-4.1 skills → GPT-5.1: 23.1% (below GPT-5.1's self-produced 25.8%)
- Skills vs. memory/knowledge: Natural-language skills are interpretable, plug-and-play transferable, and require no parameter access — friendly to closed-source APIs, unlike parameter-internalization approaches.
- Minimal viable feedback: With no ground truth, binary pass/fail verdicts plus adversarial pressure form a weak but effective feedback signal driving skill evolution.
- Diminishing gains for stronger models: Absolute gains shrink with model strength (+5.4 → +4.6 → +3.2), implying strong models already internalize general strategies and the bottleneck shifts to precise knowledge extraction.
- Compute cost: 5 iterations × 5 tasks × multiple agents per context means heavy API usage.
- Judge dependence: Binary verdicts rely on GPT-5.1 quality; judge bias distorts feedback.
- Context assumption: All required knowledge must be in the given context; no external retrieval.
- Static skills: 51.1% of tasks are sequential, but skills are pre-set and don't adapt to dialogue history.
- Collapse risk: Cross-Time Replay mitigates but does not eliminate adversarial collapse.
- Si, S. et al. (2026). *From Context to Skills: Can Language Models Learn from Context Skillfully?* arXiv:2604.27660.
- Dou, S. et al. (2026). CL-Bench: A benchmark for context learning. arXiv:2602.03587.
- Yang, Y. et al. (2026). AutoSkill: Experience-driven lifelong learning via skill self-evolution. arXiv:2603.01145.
- Zhang, H. et al. (2026). CoEvoSkills: Self-evolving agent skills via co-evolutionary verification. arXiv:2604.01687.
The five roles
| Role | Responsibility | Evolution | |------|----------------|-----------| | Challenger | Generates probing tasks + rubrics from context | Evolves harder question strategies | | Reasoner | Solves tasks using its current skill set | Evolves problem-solving skills | | Judge | Neutral binary pass/fail verdicts | Fixed (GPT-5.1), does not evolve | | Proposer (one per side) | Diagnoses failure/success cases, proposes improvements | Failure side feeds Reasoner; success side feeds Challenger | | Generator (one per side) | Converts diagnoses into concrete skill edits | Adds/deletes/merges skill entries |
Workflow per iteration: the Challenger generates M tasks with rubrics, the Reasoner attempts them, the Judge splits results into failure and success sets. Failures drive Reasoner-side skill updates; successes drive the Challenger to produce harder tasks. Crucially, the Challenger also evolves — without adversarial pressure, tasks become too easy and knowledge gaps stay hidden. Both sides evolve through failure-driven text editing, never seeing each other's skills.
Cross-Time Replay: preventing adversarial collapse
A classic self-play risk is collapse: the Challenger drifts toward extreme, unrepresentative tasks while Reasoner skills over-specialize. Cross-Time Replay instead selects the most balanced skill set across all historical iterations using probes that require no external supervision:
Selection maximizes the product (with Laplace smoothing): argmax_i (rho_h(i) * rho_e(i)), so a skill set that sacrifices easy tasks to solve hard ones gets penalized. Empirically, fixed-iteration skill performance declines monotonically (Iter-1: 15.9% → Iter-5: 14.7%), while Cross-Time Replay achieves 16.5%, beating all fixed iterations.
Experiments on CL-bench
Dataset: 500 complex contexts, 1,899 tasks, 31,607 validation rubrics across four categories (Domain Knowledge Reasoning, Rule System Application, Procedural Task Execution, Empirical Discovery & Simulation). Average 10.4K tokens, up to 65K; 51.1% of tasks are sequential/multi-turn.
Main results
| Model | Baseline | Ctx2Skill | Gain | |-------|----------|-----------|------| | GPT-4.1 | 11.1% | 16.5% | +5.4% | | GPT-5.1 | 21.1% | 25.8% | +4.6% | | GPT-5.2 | 18.2% | 21.4% | +3.2% |
Compared baselines: single-round prompting gives GPT-4.1 only +1.2%; AutoSkill4Doc (windowed extraction) gives +2.1%. Notably, GPT-4.1 + Ctx2Skill (16.5%) surpasses skill-less Gemini 3 Pro (15.8%) — skills can bridge model capability gaps.
Skill quality
Judged by GPT-4.1 on five dimensions, Ctx2Skill beats both baselines across conciseness (85.2), faithfulness (84.8), clarity (96.2), effectiveness (90.5), and reusability (92.5) — the skills are clearer, human-readable, editable, and reusable.
Ablations (GPT-4.1, 16.5% full)
Skill transferability
This asymmetry suggests skill discovery is itself an emergent capability of stronger models.
Analysis highlights
Limitations
Conclusion
Ctx2Skill's contribution goes beyond performance: it offers a general unsupervised self-evolution framework built on multi-agent adversarial co-evolution, failure-driven diagnose-then-fix loops, cross-time replay for stability, and plug-and-play natural-language skills. It marks a key step from manual prompt engineering toward automated, skill-based context learning, with applications in enterprise knowledge bases, research assistants, and education.