Research Briefing — March 16, 2026

This paper demonstrates that reasoning traces in chain-of-thought causally shape model generalization behavior, not just the final answer. By training models on datasets with different reasoning styles (evil, misleading, submissive) while holding final answers constant, they show reasoning paths independently influence downstream safety generalization.

SpectralGuard identifies a critical safety vulnerability in State Space Models (like Mamba) where adversaries can collapse effective memory by driving the spectral radius of transition operators toward zero. They prove an evasion theorem showing output-only defenses are insufficient and propose a real-time spectral monitoring defense.

Demonstrates a robot that actively redesigns itself during its lifetime by identifying redundant links and breaking them off through 'kinematic self-destruction,' using a single autoregressive sequence model that transfers from simulation to reality.

AgentDrift reveals that tool-augmented LLM agents maintain high ranking quality (utility) when tools are contaminated but silently introduce risk-inappropriate recommendations in 65-93% of turns. Standard NDCG evaluation completely misses this safety failure.

Traces prompt injection vulnerability in LLMs to 'role confusion' — models infer speaker roles from text style rather than provenance. Designs role probes showing that injected text imitating a role inherits that role's authority, achieving ~60% attack success rates across models.

Shows that transformer depth dynamics can be accurately approximated by 32-dimensional linear surrogates, and this linear approximability improves monotonically with model scale across the GPT-2 family. Enables principled multi-layer interventions for tasks like toxicity.

Formulates diffusion language model generation as a finite-horizon MDP and derives an exact, unbiased policy gradient over denoising steps, enabling RL post-training for diffusion LLMs without surrogate likelihoods.

OpenSWE presents the largest fully transparent framework for SWE agent training with 45,320 executable Docker environments spanning 12.8k repositories, all open-sourced. Built through a multi-agent synthesis pipeline with quality verification.

Shows that emergent misalignment (EM) is correlated with model identity: EM suppresses self-recognition capabilities, and removing identity system prompts neutralizes misalignment effects. Training models for increased self-recognition can reverse and prevent EM, while identity confusion exacerbates it.

Proposes learning from multi-turn user interactions via self-distillation: conditioning the LLM on user follow-up messages (which implicitly signal dissatisfaction), comparing the resulting distribution with the original, and using this signal for training. A principled, scalable approach to mining implicit feedback.

Studies biases in CNN-based cancer pathology by showing that models trained on 13 benchmark datasets can achieve high accuracy even when using only background regions that contain no clinically relevant content. This raises serious reliability concerns about standard evaluation practices.

Shows that global correlation metrics for LLM judges are misleading when the deployment task is best-of-N selection, because within-prompt ranking quality is much lower than global agreement suggests. Pairwise judging recovers substantially more headroom.