Research Briefing — April 22, 2026

Evaluates LLM-based scientific agents across 8 domains with 25,000+ agent runs, finding that agents produce results without adhering to epistemic norms of scientific reasoning. The base model accounts for 41.4% of performance variance, dominating the scaffold.

Proposes a test-time scaling framework for coding agents that converts rollout trajectories into structured summaries preserving hypotheses, errors, and partial progress, enabling effective selection and reuse across attempts. Features authors from CMU, Meta, and Princeton.

Presents OmniMouse, a multi-modal multi-task brain model trained on 3.1 million neurons from 73 mice (150B neural tokens). Demonstrates scaling laws for neural data modeling, achieving state-of-the-art across neural prediction, behavioral decoding, and neural forecasting tasks.

Shows that across 12 open-weight LLMs, the same small set of attention heads carries a 'this is wrong' signal during both sycophancy and factual lying, revealing a shared circuit. Silencing these heads flips sycophantic behavior while preserving factual accuracy.

Identifies the 'Trust Gap' in tool-using AI agents - they're evaluated for capability but not skepticism. Introduces Adversarial Environmental Injection (AEI) and POTEMKIN benchmark where tool outputs are poisoned to deceive agents.

Shows that safety failures in large reasoning models stem from the reasoning structure itself, and proposes AltTrain, a simple SFT method using only 1K examples to alter reasoning structure for safety alignment without complex RL.

Studies using pretrained Sparse Autoencoders (SAEs) inserted into transformer residual streams at inference time as a defense against jailbreak attacks, achieving up to 5x reduction in attack success rate. Tests across four model families (Gemma, LLaMA, Mistral, Qwen) and multiple attack types without modifying model weights. Reveals a monotonic relationship between SAE width and robustness.

EVPO introduces explained variance as a criterion for adaptively choosing between critic-based (PPO) and critic-free (GRPO) baselines in RL for LLM post-training. Unifies PPO and GRPO as extremes of a Kalman gain and shows a learned critic can increase variance in sparse-reward settings.

Presents an alternative view of generalization that separates geometric from probabilistic assumptions, using sensitivity analysis rather than i.i.d. assumptions. Authors include Benjamin Recht, a prominent ML theorist.

NSF workshop report identifying three AI capability gaps (physical world interaction, brittle learning, energy inefficiency) and mapping neuroscience principles that could address each. Features prominent authors (Zador, Sejnowski, etc.).

Demonstrates that harmful intent is geometrically recoverable from LLM residual streams as a linear direction in most layers (AUROC 0.98) across 12 models spanning four architectural families and three alignment variants. Angular deviation succeeds where linear projection fails.

Introduces SimpleTES, a general framework for scaling evaluation-driven scientific discovery loops by strategically combining test-time compute with evaluation feedback. Addresses how to principally scale trial-and-error discovery with LLMs.