English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Ctx2Skill: Multi-Agent Self-Play for Self-Evolving Context Skills in LLMs

Forum topic · 小凯 · 2026-05-16

Summary

Ctx2Skill (arXiv:2604.27660, Tsinghua, DeepLang AI, UIUC, Fudan, CUHK) is a framework that lets large language models autonomously extract reusable skills from complex, previously unseen contexts without human labels or ground-truth feedback. A Challenger agent generates probe tasks with rubrics, a Reasoner solves them, and a fixed Judge gives binary pass/fail verdicts. Proposer-Generator pairs on each side perform diagnose-then-fix text edits on natural-language skill sets in response to failures and successes, producing co-evolution under strict adversarial pressure. Cross-Time Replay selects the most balanced skill set across all iterations using hard and easy probes to prevent adversarial collapse. On CL-Bench (500 contexts, 1,899 tasks, 31,607 rubrics), GPT-4.1 rises from 11.1% to 16.5%, GPT-5.1 from 21.1% to 25.8%, and GPT-5.2 from 18.2% to 21.4%. Skills are human-readable, transferable across models, and outperform prompting and AutoSkill4Doc baselines.

Overview

Paper: *From Context to Skills: Can Language Models Learn from Context Skillfully?* (Ctx2Skill) 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 Published: arXiv:2604.27660 (v1 2026-04-30, v2 2026-05-03) Code: https://github.com/S1s-Z/Ctx2Skill

---

The Problem: Context Learning, Not Just Long-Context Reading

Modern LLMs struggle when given a novel, complex context outside their pre-training distribution (e.g., an unfamiliar technical manual, experiment data, domain paper). The challenge is context learning: inducing new knowledge from the context and applying it to solve tasks, rather than simple retrieval or reading comprehension.

Two existing approaches hit hard limits:

  • Manual annotation is prohibitively expensive for long, dense documents.
  • Automated construction lacks feedback: unlike code or math, context-learning tasks have no ground truth to evaluate whether extracted skills are complete or faithful.
  • > "Given only the context, there is no external feedback signal to tell whether a generated skill is useful or whether critical knowledge has been omitted."

    ---

    Core Idea: Multi-Agent Self-Play Skill Evolution

    Instead of relying on human labels, Ctx2Skill lets the model play against itself: one agent generates tasks, another solves them, a judge scores the results, and losing side reflects and upgrades its skills.

    Five Roles

    | Role | Responsibility | Evolution | |------|----------------|-----------| | Challenger | Generates probe tasks and rubrics from the context | Evolves task-generation strategy | | Reasoner | Solves tasks with current skill set | Evolves solving skills | | Judge | Binary pass/fail verdict | Fixed (GPT-5.1); no evolution | | Proposer (one per side) | Diagnoses failures/successes, proposes high-level improvements | Updated each round | | Generator (one per side) | Translates proposals into concrete skill edits (add/delete/merge) | Updated each round |

    Workflow per Iteration

    1. Challenger (skill set S^C_{i-1}) produces M tasks with rubrics. 2. Reasoner (skill set S^R_{i-1}) attempts to solve them. 3. Judge splits outcomes into failure set F_i and success set P_i. 4. Failure set → Reasoner Proposer diagnoses missing knowledge → Reasoner Generator updates S^R_i. 5. Success set → Challenger Proposer diagnoses how to make tasks harder → Challenger Generator updates S^C_i. 6. Iterate under strict adversarial isolation (sides cannot peek at each other's skills).

    Key design: The Challenger also evolves. If it stayed static, tasks would become too easy and the Reasoner would stop exposing knowledge gaps.

    ---

    Cross-Time Replay: Preventing Adversarial Collapse

    Risk

    Pure self-play tends toward adversarial collapse: the Challenger invents increasingly pathological tasks; the Reasoner's skills overfit to those edge cases; redundancy accumulates; generalization collapses.

    Solution

    Cross-Time Replay does not simply take the final-iteration skill set. It searches all historical iterations and picks the most balanced one.

    Probe sets (built without external supervision):

  • Hard probe Q^h: per iteration, pick the failing tasks that passed the fewest rubrics (hardest failures).
  • Easy probe Q^e: per iteration, pick the successful tasks that passed the fewest rubrics (easiest successes).
  • Selection criterion (product form with Laplace smoothing):

    S^R_* = argmax_i ( ρ^h(i) · ρ^e(i) )

    where ρ^h(i) is the solve rate of iteration-i skills on the hard probe, and ρ^e(i) on the easy probe. The product punishes any skill set that sacrifices easy tasks to brute-force hard ones, or vice versa.

    Empirical validation: Fixed-iteration skill performance drops monotonically (Iter-1: 15.9% → Iter-5: 14.7%). Cross-Time Replay reaches 16.5%, surpassing every fixed iteration.

    ---

    Experiments

    Benchmark: CL-Bench

  • 500 complex contexts
  • 1,899 tasks
  • 31,607 verification rubrics
  • Four categories: Domain Knowledge Reasoning, Rule System Application, Procedural Task Execution, Empirical Discovery & Simulation
  • Average context length 10.4K tokens, max 65K tokens
  • 51.1% of tasks are multi-turn sequential (later steps depend on earlier solutions)
  • Main Results

    | Model | Baseline | + Ctx2Skill | Δ | |-------|----------|-------------|---| | 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 with alternatives:

  • Prompting (single-shot skill generation): GPT-4.1 only +1.2%; some categories regress.
  • AutoSkill4Doc (windowed extraction): GPT-4.1 +2.1%, far below Ctx2Skill's +5.4%.
  • Headline finding: GPT-4.1 + Ctx2Skill (16.5%) outperforms skill-less Gemini 3 Pro (15.8%). Skills can bridge raw capability gaps between models.

    Skill Quality (GPT-4.1 as judge)

    | Dimension | Prompting | AutoSkill4Doc | Ctx2Skill | |-----------|-----------|---------------|-----------| | Conciseness | 81.2 | 81.3 | 85.2 | | Faithfulness | 79.7 | 81.4 | 84.8 | | Clarity | 80.0 | 92.4 | 96.2 | | Effectiveness | 83.3 | 88.7 | 90.5 | | Reusability | 84.7 | 87.2 | 92.5 |

    Self-play skills are not just more effective; they are also more structured, human-readable, editable, and reusable.

    Ablations (GPT-4.1 overall)

    | Variant | Score | Δ vs full | |---------|-------|-----------| | Full Ctx2Skill | 16.5% | — | | – w/o Challenger evolution | 13.8% | −2.7% (largest drop) | | – w/o Cross-Time Replay | 14.7% | −1.8% | | – w/o Hard Probe | 15.2% | −1.3% | | – w/o Easy Probe | 15.7% | −0.8% | | – w/o Laplace Smoothing | 15.5% | −1.0% | | – Proposer+Generator merged | 15.9% | −0.6% |

    Skill Transferability

    | Skill source | Reasoner model | Result | |--------------|----------------|--------| | GPT-5.1 skills → GPT-4.1 | GPT-4.1 | 16.1% (≈ GPT-4.1's own 16.5%) | | GPT-4.1 skills → GPT-5.1 | GPT-5.1 | 23.1% (< GPT-5.1's own 25.8%) |

    Asymmetry: Strong-model skills transfer well to weaker models; weak-model skills help strong models only marginally. This hints that skill discovery itself may be an emergent capability.

    ---

    Analytical Insights

    Why natural-language skills, not memory or parameters?

    1. Interpretability: skills are human-readable Markdown, auditable and editable. 2. Transferability: plug-and-play across models. 3. No parameter access required: friendly to closed-source APIs.

    This is more practical than parameter-internalization approaches like SKILL0/SkillRL in commercial API settings.

    Meta-stability of self-play

    The system forms a self-correcting loop:

  • Challenger too strong → Reasoner keeps failing → Challenger-side evolution (too many successes) → Challenger rebalances.
  • Challenger too weak → Reasoner keeps succeeding → Reasoner-side evolution stalls → Challenger-side evolution (too many successes) forces harder tasks.
  • Cross-Time Replay keeps either side from drifting to extremes.

    Minimal viable feedback

    Without ground truth, binary pass/fail under adversarial pressure is a weak but sufficient learning signal. The system does not need to know the correct answer; it only needs to know whether the answer cleared all rubrics, and changes in pass-rate drive skill improvement.

    Why smaller relative gains for stronger models

    GPT-4.1 sees +48.6% relative lift, GPT-5.1 +21.8%, GPT-5.2 +17.6%. Stronger models already internalize general problem-solving strategies, so the bottleneck shifts from "missing strategies" to "precise knowledge extraction." Even so, +4.6% on a 21% baseline is significant.

    Comparison with LongNAP

    | Dimension | LongNAP (2603.05923) | Ctx2Skill (2604.27660) | |----------|----------------------|------------------------| | Data source | Passive mobile screenshots + clicks | Active, given context | | Learning goal | Predict next user action | Extract reusable skills for tasks | | Feedback | Delayed (compare with later user action) | Binary Judge (pass/fail) | | Training | Policy Gradient + LLM-as-judge | Self-play + text editing | | Core innovation | Retrieving user history | Adversarial evolution + cross-time replay |

    Both point to the same direction: let AI self-evolve through environmental interaction rather than human annotation.

    ---

    Limitations and Outlook

    | Limitation | Notes | |------------|-------| | Compute cost | 5 rounds × 5 tasks × multiple agents per context → heavy API usage; larger N/M not explored due to budget | | Judge dependency | Binary verdicts inherit GPT-5.1 biases | | Context assumption | All required knowledge assumed inside the given context; no external retrieval | | Multi-turn dependency | 51.1% sequential tasks need prior solutions, but skills are static and not dynamically adapted to dialog history | | Collapse risk | Cross-Time Replay mitigates but does not eliminate adversarial collapse |

    Application prospects: enterprise knowledge bases (auto-extract operating procedures), research assistants (induce methodology from papers and data), education (extract problem-solving strategies from textbooks).

    ---

    Takeaways

    Ctx2Skill contributes more than a benchmark number. It offers a general unsupervised self-evolution framework:

    1. Multi-agent adversarial co-evolution creates sustained optimization pressure. 2. Failure-driven diagnose-then-fix (Proposer-Generator) beats single-shot generation. 3. Cross-Time Replay picks the best historical skill set and prevents collapse. 4. Plug-and-play natural-language skills transfer across models and work with closed-source APIs.

    > "We hope Ctx2Skill provides a practical and scalable paradigm for equipping language models with the ability to learn skillfully from complex, previously unseen contexts."

    Context learning moves from hand-engineering to automated evolution.

    ---

    References

  • 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.

Tags

#llm#context-learning#multi-agent#self-play#skill-extraction#adversarial-training#prompting#benchmark

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620123