If AI agents learn from each other instead of one fixed teacher, can they succeed from messier starting points?
How does cross-agent supervision expand the set of convergent initial conditions?
This explores whether having other agents (partners, peers, or simulated environments) supply the training signal lets learning succeed from a wider range of starting points, rather than only from a narrow, well-prepared one. The corpus never measures 'convergent initial conditions' directly, so this is assembled from adjacent findings.
This explores whether having other agents (partners, peers, or simulated environments) supply the training signal lets learning succeed from a wider range of starting points, rather than only from a narrow, well-prepared one. No note in the collection tests that directly, but several point the same way. Training from a single fixed source of supervision seems to shrink the set of starting points that work, and supervision that comes from other agents, or is generated relative to the learner's current behavior, seems to grow it.
The clearest case for a single fixed supervisor is what it leaves out. Agents trained on static expert demonstrations never see their own failures, so their competence is capped by what the curator imagined rather than by the agent's capacity (Can agents learn beyond what their training data shows?). Only starting points that already resemble the expert's world can get anywhere. The alternatives loosen that dependence. Agents that treat the future states produced by their own actions as supervision matched expert-dependent baselines with half the data, and gave better warm-starts for later RL (Can agents learn from their own actions without external rewards?). A learned language world model can stand in for the environment and scale agent training beyond what real environments allow (Can language models learn to simulate agent environments?). In both, the feedback is produced from wherever the agent currently is.
The strongest cross-agent evidence is about cooperation. Sequence-model agents trained against a diverse set of co-players learn in-context best-response strategies, and these settle into cooperation without hardcoded assumptions about partners or separate timescales (Can agents learn cooperation by adapting to diverse partners?). The other agents act as supervision. Because every agent is vulnerable to being exploited by the others, pressure toward cooperation appears from many different partner mixes, instead of only from a carefully designed starting setup. A related, weaker version is agreement across a model's own samples. Majority-vote consensus can replace ground-truth labels in self-distillation, and it trains only on rollouts where the samples disagree (Can a model's own consensus replace ground truth labels?). That is a same-model version of the idea, not true cross-agent supervision.
Two notes qualify the picture. Long-horizon success is predicted more by persistence in repeated benchmark-edit cycles than by how good the first attempt was (What predicts success in ultra-long-horizon agent tasks?). That fits the idea that a live feedback loop matters more than the starting point, though it concerns feedback loops generally, not other agents. Cross-agent supervision also only helps when the other agents are genuinely separate. LLMs look socially competent when one model plays every party, but fail systematically once agents hold private information (Why do LLMs fail when simulating agents with private information?). A supervisor that knows everything can make a bad starting point look fine.
The working hypothesis is that supervision generated from the learner's own current behavior, or from partners who differ from it, stays informative from many starting points. A fixed expert dataset only helps starts that already resemble it. The gap is that nobody in this collection has mapped which starting points converge and which don't, so the widening is inferred from results, not measured.
Sources 7 notes
Agents trained on static expert datasets cannot learn from their own failures or generalize beyond demonstrated scenarios because they never interact with environments during training. Competence is capped by what curators imagined, not by agent capacity.
Research across eight environments shows that agents can use future states from their own actions as supervision without external rewards, matching expert-dependent baselines with half the data and providing superior warm-starts for subsequent RL training.
Qwen-AgentWorld demonstrates that native language world models trained via next-state prediction on 10M+ trajectories outperform real-environment training on three benchmarks and transfer across seven domains, positioning next-state prediction as a foundation objective for agents.
Sequence model agents trained against diverse co-players develop in-context best-response strategies that naturally resolve into cooperation. Mutual vulnerability to exploitation creates pressure that drives cooperative mutual adaptation without hardcoded assumptions or timescale separation.
Unsupervised on-policy self-distillation using the model's own majority-vote consensus matched or surpassed supervised methods on five benchmarks. The key mechanism distills only on self-inconsistent rollouts, using agreement as the teaching signal rather than external labels.
Show all 7 sources
Across 17 frontier models on 36 expert-curated optimization tasks, repeated benchmark-edit-incorporate cycles within a wall-clock budget proved the dominant success predictor. Most models terminated early or burned budget unproductively; Claude Opus 4.6 stood out as persistent.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Agent Learning via Early Experience
- Is this the real life? Is this just fantasy? The Misleading Success of Simulating Social Interactions With LLMs
- SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
- Self-distillation Enables Continual Learning
- Qwen-AgentWorld: Language World Models for General Agents
- On-Policy Self-Distillation without Any Supervision
- Multi-agent cooperation through in-context co-player inference
- Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks