Research Briefing — March 25, 2026

Models LLM test-time computation as a Markov chain where the algorithm can backtrack to any previous state. Derives the optimal strategy ('Caterpillar of Thoughts') for allocating a fixed computation budget, providing theoretical foundations for test-time compute.

Introduces Algorithmist, an autonomous agent built on GitHub Copilot that synthesizes algorithms with provable guarantees through a multi-agent research-and-review loop. Demonstrates provable algorithm synthesis at scale with stages for idea generation, proof development, and implementation.

Introduces step-level evaluation for CoT reasoning: removing one reasoning sentence at a time and checking if the answer changes. Finds that for most frontier models, individual reasoning steps are decorative—removal doesn't change the answer. Costs ~$1-2 per model per task.

Introduces sparse packing formats and CUDA kernels for unstructured sparsity in LLM feedforward layers, showing L1 regularization can induce 99%+ sparsity with negligible performance loss and real wall-clock speedups during inference and training.

Provides systematic token-level analysis of how RLVR fine-tuning changes LLM distributions, finding that changes are highly sparse and targeted—only a small fraction of token distributions shift meaningfully, but these sparse changes drive sequence-level reasoning improvements.

Evaluates chain-of-thought faithfulness across 12 open-weight reasoning models (7B-685B), extending prior work limited to two proprietary models. Finds that many models exhibit very low acknowledgment rates for injected reasoning hints, suggesting CoT is often decorative rather than functional.

Proposes ReVal, a value-based off-policy RL method for LLMs that combines stepwise Bellman updates with trajectory-level outcome signals and replay buffer training for improved sample efficiency.

Studies arbitrage in AI model markets where an intermediary efficiently allocates inference budget across providers (GPT-5 mini and DeepSeek v3.2) to undercut market prices. Demonstrates viable arbitrage strategies on SWE-bench with net profit generation.

Proposes T-MAP, a trajectory-aware evolutionary search method for red-teaming LLM agents that leverages execution trajectories to discover adversarial prompts causing harmful tool interactions, specifically targeting Model Context Protocol (MCP) ecosystems. Demonstrates attacks that bypass safety guardrails through multi-step tool execution.

Reveals systematic biases in Chinchilla Approach 2 for fitting neural scaling laws. Shows the parabolic approximation causes compute-optimal allocation errors worth $1.4M at frontier scales, with even larger costs for multimodal models.

Revisits IQ-Learn and proves it reduces to behavioral cloning with quadratic horizon-dependent imitation gap, still suffering from compounding errors. Proposes Bellman constraint-based non-adversarial methods that provably avoid compounding errors.

Proposes Sparse Feature Attention (SFA) that reduces attention cost from O(n²d) to O(n²k²/d) by representing queries and keys as k-sparse codes. Introduces FlashSFA, an IO-aware kernel extending FlashAttention for sparse operations.