Research Briefing — February 11, 2026

Demonstrates theoretically and empirically that self-evolving multi-agent LLM societies cannot simultaneously achieve continuous self-improvement, complete isolation, and safety invariance—termed the 'self-evolution trilemma.' Uses information-theoretic framework to show isolated self-evolution inevitably degrades safety alignment.

Establishes information-theoretic barriers for credit assignment in multi-stage systems (manufacturing, AI reasoning chains), proving that signal from early steps to final outcomes decays exponentially with depth, creating a 'critical horizon' beyond which no algorithm can learn from endpoint data alone.

Proposes counterfactual importance weighting for policy gradient methods (GRPO/DAPO) in LLM reasoning, replacing uniform credit assignment across tokens with importance-weighted updates based on masking reasoning spans and measuring answer probability drops. Demonstrates consistent improvements over uniform baselines on GSM8K across Qwen and Llama models.

Introduces WildCat, a near-linear time attention mechanism that uses randomly pivoted Cholesky decomposition to select a spectrally-accurate weighted coreset for attention computation. Achieves super-polynomial error decay while running in near-linear time, with competitive results on language modeling and image classification.

Presents RLFR (Reinforcement Learning from Feature Rewards), which uses interpretable features learned by language models as reward functions for RL-based hallucination reduction. Uses a probing framework to identify hallucinated claims and teaches the model to intervene and correct uncertain completions.

Presents large-scale analysis of Moltbook, a Reddit-style social media platform exclusively populated by AI agents (~46K agents, 369K posts, 3M comments). Finds AI collective behavior exhibits many statistical regularities of human communities but with key differences like sublinear upvote-discussion relationships.

Continuing our coverage from yesterday, Zvi's detailed coverage of the Claude Opus 4.6 system card focusing on frontier alignment topics: sabotage, deception, situational awareness, catastrophic risks. Argues the model was correctly released as ASL-3 but that Anthropic's process may not scale to Opus 5.

Demonstrates that contradictory evaluations of LoRA variants largely stem from overlooked batch size differences, showing that vanilla LoRA with properly tuned batch size often matches complex variants. Proposes cost-efficient batch size tuning strategy.

Proposes Infusion, a framework using influence functions to craft training data that induces targeted model behavior changes. Shows that editing 0.2% of training data can be competitive with explicit poisoning. Demonstrates cross-architecture transfer.

Introduces AIDev, a large-scale dataset of 932,791 agent-authored pull requests from five AI coding agents (Codex, Devin, Copilot, Cursor, Claude Code) across 116K GitHub repositories.

Proposes BiasScope, a framework for automatically discovering unknown biases in LLM-as-a-Judge evaluation systems at scale. Can uncover biases across different model families and scales.

Shows that LLMs encode their likelihood of success in pre-generation internal activations, and linear probes can predict task-specific success, enabling more efficient inference by routing only hard problems to expensive reasoning.