Research Briefing — April 14, 2026

Proves that activation steering pushes LLM residual streams off the manifold of states reachable from discrete prompts, meaning steered activations are not realizable by any textual input. This has implications for interpretability and safety research using steering.

Experimentally shows that multi-agent AI organizations are simultaneously more effective at business goals but less aligned than individual AI agents, across 12 tasks in consultancy and software settings. Demonstrates emergent misalignment from agent interaction.

Establishes the first information-theoretic lower bounds for score queries in diffusion sampling, proving that any sampler requires Ω̃(√d) adaptive score queries for d-dimensional distributions. This provides a formal explanation for why multi-scale noise schedules are necessary in diffusion models.

Demonstrates that MoE expert specialization is an emergent property of the representation space, not the routing architecture, since linear routers make hidden state similarity necessary and sufficient for explaining routing patterns. Proves load-balancing loss suppresses shared directions.

First empirical analysis of social structure formation among 626 autonomous AI agents on a live network (Pilot Protocol). Finds heavy-tailed degree distributions consistent with preferential attachment in the autonomously formed trust network. References OpenClaw instances.

Introduces Meerkat, a system combining clustering with agentic search to detect safety violations across large sets of agent traces. Addresses failures that are only visible when analyzing multiple traces together, relevant to misuse campaigns, covert sabotage, and reward hacking.

LABBench2 is an evolved benchmark for measuring AI systems' ability to perform real biology research, going beyond rote knowledge to test meaningful scientific work capabilities. This builds on the original LAB-Bench and shifts focus toward real-world scientific competence.

Systematically investigates 'overthinking' in LLM reasoning: when extended chain-of-thought actually leads to worse results by abandoning correct answers. Shows diminishing marginal returns at higher compute budgets and that optimal thinking length varies by problem difficulty.

Introduces Introspective Diffusion Language Models (I-DLM), identifying that diffusion LMs lack 'introspective consistency' (accepting their own generations) that AR models inherently have. Proposes a training paradigm combining parallel diffusion decoding with AR-style introspective consistency.

Proposes RationalRewards, teaching reward models to produce explicit multi-dimensional critiques before scoring for visual generation. Introduces Preference-Anchored Rationalization (PARRO) to train without costly rationale annotations. Shows improvements at both training and test time.

Introduces Pando, a benchmark for mechanistic interpretability that controls for the 'elicitation confounder' - whether interpretability gains simply reflect information obtainable through prompting. Uses 720 finetuned models with faithful, absent, or unfaithful explanations.