Research Briefing — March 24, 2026

Investigates whether LLMs causally use internal confidence estimates to drive abstention behavior through a four-phase paradigm, finding that confidence is the dominant predictor of behavior with effect sizes an order of magnitude larger than knowledge retrieval access. Authors include DeepMind researchers.

Reports three findings challenging multimodal AI assumptions: models generate detailed descriptions for images never provided ('mirage reasoning'), achieve high benchmark scores without image input, and show significant visual-textual decoupling. Fei-Fei Li is a co-author.

Proves that LLMs report highest confidence when fabricating, with self-reported confidence inversely correlating with accuracy. Formally proves this is an observational limitation: text-only monitoring cannot reliably distinguish honest outputs from fabrications.

Introduces latent lookahead training that enables transformer models to perform additional forward passes in latent space before committing to the next token, allowing the model to 'think' and allocate more compute to difficult tokens.

Introduces 'Thought Injection' to test whether LLM reasoning traces causally influence outputs and whether models acknowledge this influence. Finds injected hints reliably alter outputs but models refuse to disclose the influence >90% of the time.

Introduces LongCat-Flash-Prover, a 560B MoE model for formal theorem proving in Lean4, using agentic tool-integrated reasoning with a Hierarchical Importance Sampling Policy Optimization algorithm.

Investigates whether independently trained LLMs converge to geometrically compatible latent representations, learning linear projections to map teacher activations into student coordinate systems. Shows Ridge projection consistently works across 20 heterogeneous teacher-student pairings.

Maps the compression sensitivity landscape across five transformer architectures (117M-8B params), finding that early-layer MLP up-projections are catastrophically sensitive while value projections compress nearly for free. Uses Lyapunov stability theory to explain why residual connections contract compression errors, providing formally verified bounds.

Introduces a depth-recurrent Transformer that decouples computational depth from parameter count by iteratively applying shared-weight blocks, enabling variable-depth reasoning at inference time. Uses silent thinking objectives and stabilization mechanisms for 20+ recurrence steps.

Presents a systematic empirical study of RL for long-horizon tool-using LLM agents using TravelPlanner, decomposing the design space along 5 axes (reward shaping, model scaling, data composition, algorithm selection, environmental stability) and yielding 7 key takeaways about scale-dependent design choices.