AI systems that learn to improve without human data now occupy a central place in research. This month, two topics related to this approach are experiencing unprecedented growth: multi-agent systems and negotiation (172 papers over four weeks, compared to 53 the previous month) and information retrieval (149 papers, compared to 47). These two fields share a common technique, on-policy self-distillation (OPSD), where models refine their responses by training on their own trials rather than on examples provided by humans.
OPSD is not new, but its massive adoption in concrete applications marks a turning point. Recent work no longer merely touts its promises: it tests its limits. Three questions recur in this month’s titles:
- How to prevent the model from misleading itself with privileged information (hidden data that biases its learning)? 29 papers address this, 19 more than in July.
- How to stabilize training when rewards are scarce or noisy? 23 papers explore the use of privileged information to guide distillation, and 20 others focus on token-level supervision (a learning signal refined at the level of each word or symbol).
- How to trace a response back to its source to verify its robustness? The phrases « evidence supports » (22 papers) and « local evidence » (17 papers) appear in work assessing the reliability of step-by-step reasoning.
A few recent examples:
- Rethinking Privileged Information in On-Policy Self-Distillation
- Beyond Teacher Likelihood: Group-Calibrated On-Policy Distillation for Long-Context Reasoning
- Trust Is Not Enough: Influence Calibration for On-Policy Self-Distillation in Agentic RL
- RuleWeaver: Benchmarking Rule-Centered Scenario Reasoning for Large Language Models
- Neuro-symbolic PRM: Enhancing Scientific Reasoning via Structured Traces and Symbolic Verification
This trend is accompanied by a marked decline in traditional approaches. Generative adversarial networks (268 papers, -23% in one month) and advanced neural networks (135 papers, -24%) are losing ground, as is explainability (107 papers, -38%). Research now seems to favor systems capable of improving on their own, even at the cost of some transparency.