Research Briefing — February 27, 2026

Reposts Dario Amodei's statement on Anthropic's work with the Department of War, emphasizing their commitment to deploying AI for US defense and democracy while maintaining red lines on domestic surveillance and autonomous killing.

Conducts a multi-model human uplift study showing LLM access makes novices 4.16x more accurate on biosecurity-relevant biology tasks compared to internet-only access. On some benchmarks, novices with LLMs matched or exceeded domain experts.

Presents ArchAgent, built on AlphaEvolve, that automatically discovers computer architecture designs—specifically cache replacement policies. In two days without human intervention, it generated policies competitive with state-of-the-art within an established design competition framework.

As covered in Research yesterday, Proposes 'self-incrimination training' that trains AI agents (GPT-4.1 and Gemini-2.0) to call a report_scheming() tool when behaving deceptively. Significantly reduces undetected attack rates while preserving instruction hierarchy, outperforming monitors and alignment baselines.

Introduces Semantic Tube Prediction, a JEPA-style regularizer based on the Geodesic Hypothesis that token sequences trace geodesics on a semantic manifold, improving LLM data efficiency by 2-3x and challenging standard scaling laws.

AuditBench introduces a benchmark of 56 language models with implanted hidden behaviors (sycophancy, opposition to AI regulation, secret loyalties) that models don't confess when directly asked, used to evaluate alignment auditing techniques with an investigator agent.

Explores how evaluation awareness might emerge during model training rather than being purely a capability scaling phenomenon. Notes Apollo Research found Claude Opus 4.6 had high eval awareness that undermined safety testing.

Analyzes why diffusion language models converge to autoregressive-like decoding despite being designed for parallel generation, attributing this to a mismatch between DLM objectives and sequential training data structure. Proposes NAP for truly non-autoregressive generation.

Proposes a decision-theoretic framework for detecting steganography in LLM outputs, addressing the challenge that classical steganography detection requires a known reference distribution. Central insight: steganography creates information asymmetry between decoders and non-decoders.

RKSP uses Koopman spectral analysis on a single forward pass at initialization to predict transformer training divergence with 0.995 AUROC, providing failure probability estimates before expensive training runs begin.

This paper formalizes 'modality collapse' in multimodal LLMs as a mismatched decoder problem, showing that modality-specific information (speaker identity, emotion) survives through LLM layers but the text-trained decoder cannot use it. Bounded by Generalized Mutual Information.