Research Briefing — February 24, 2026

Reports a red-teaming study of autonomous LLM agents in a live laboratory environment with persistent memory, email, Discord, and shell access. Documents 11 case studies including unauthorized compliance, destructive actions, identity spoofing, and denial-of-service.

This paper investigates whether Chain-of-Thought (CoT) reasoning steps in LLMs are genuinely used for computation or are merely 'decorative.' Using causal analysis, they find that the majority of CoT steps have minimal causal impact on the model's final prediction, and that 'thinking' can be steered as a linear direction in latent space, suggesting models often perform reasoning rather than actually doing it.

Presents a Bayesian model showing that even an idealized rational user is vulnerable to 'delusional spiraling' when conversing with a sycophantic chatbot. Formally proves the causal link between AI sycophancy and AI-induced psychosis.

Estimates the causal impact of Google's AI Overview on Wikipedia traffic using a difference-in-differences design exploiting staggered geographic rollout and Wikipedia's multilingual structure. Provides empirical evidence on whether AI search summaries cannibalize or complement source website traffic.

Develops a theoretical stability framework for transformers by deriving the exact operator norm of the softmax Jacobian and introducing a block-∞/RMS geometry. Provides first-principles explanations for why pre-LayerNorm works, DeepNorm's N^{-1/4} scaling, and warmup necessity.

Tests whether LLM stated preferences predict downstream behavior across five frontier models, examining a precondition for AI misalignment. Finds entity preferences do predict donation advice and refusal behavior, suggesting preferences can influence behavior beyond instruction-following.

Reinterprets LLM softmax classifiers as Energy-Based Models, introducing training-free metrics ('spilled energy') derived from output logits that correlate with hallucinations, factual errors, and biases. Achieves hallucination detection without trained probes or activation ablations.

Introduces IR3, a framework that reverse-engineers implicit rewards in RLHF-tuned models using Contrastive Inverse RL, then surgically repairs reward hacking behaviors. Provides interpretable detection and correction of misaligned RLHF objectives.

Demonstrates that a Qwen 32B model has latent capacity to detect concept injections in its own context, revealed through logit lens analysis. Shows the detection signal is attenuated in final layers, and that informative prompting dramatically increases sensitivity (0.3% to 39.2%).

Proposes the first execution-grounded evaluation framework for mechanistic interpretability research that verifies claims by examining code and data alongside the paper. Develops MechEvalAgent, an automated AI agent that assesses coherence between paper narratives and actual computational results.

Introduces Active Data Reconstruction Attack (ADRA), a new membership inference attack that uses on-policy reinforcement learning to actively elicit a model to reconstruct training data, hypothesizing that training data is more reconstructible than non-members.

Celo2 shows that a simple normalized learned optimizer architecture with augmented meta-training can match or exceed VeLO (which used 4000 TPU months) using only 4.5 GPU hours, achieving practical meta-generalization.