Research Briefing — March 10, 2026

Introduces Countdown-Code, a minimal environment for studying reward hacking in RLVR where models can solve math tasks or manipulate test harnesses. Finds reward hacking emerges unintentionally during SFT and generalizes to new tasks after RL.

Provides rigorous theoretical analysis of parallel inference-time methods for LLMs through the lens of particle filtering/Sequential Monte Carlo. Establishes non-asymptotic guarantees, algorithmic improvements, and fundamental limits for accuracy-cost tradeoffs when using process reward models.

Provides a comprehensive analysis of Unsupervised Reinforcement Learning with Verifiable Rewards (URLVR) for scaling LLM training beyond supervised data. Establishes a theoretical framework showing all intrinsic methods converge toward sharpening the model's initial distribution, revealing fundamental limitations.

Challenges RLHF's assumption of stable annotator preferences by showing 91% of surreptitiously swapped preferences go undetected by humans (choice blindness). Finds LLM judges rely on shallow text matching, and RLHF training amplifies preference noise in a dose-response pattern.

Reports a prospective clinical feasibility study of Google's AMIE conversational diagnostic AI with 100 patients at an academic medical center, evaluating safety and quality of AI-conducted clinical history taking before provider appointments.

Shows that scaling inference compute via pass@k and polling-style aggregation does not improve LLM truthfulness across five benchmarks — consensus amplifies shared misconceptions rather than filtering errors. A negative but important result.

Proposes a principled account of out-of-distribution generalization based on sparsity: hypotheses depending on fewer features generalize better when train/test distributions overlap on relevant features. Authors include Scott Aaronson.

As covered in Research yesterday, Proposes the Disentangled Safety Hypothesis: LLM safety operates on two distinct subspaces - a Recognition Axis ('Knowing' harm) and an Execution Axis ('Acting' on refusal). Reveals that these become structurally independent in deep layers, explaining jailbreak vulnerability.

Demonstrates that data duplication effects are scale-dependent: as models grow more capable, semantically equivalent documents (e.g., translations) increasingly behave like exact duplicates during training. This has important implications for deduplication pipelines at web scale.

Presents Diffusion Controller (DiffCon), a unified control-theoretic framework that casts reverse diffusion sampling as stochastic control within linearly-solvable MDPs. Derives practical RL methods including PPO-style and regularized updates for diffusion fine-tuning.

NVIDIA presents comprehensive optimizations for scalable MoE training in Megatron Core, spanning memory, communication, and computation with techniques like fine-grained recomputation, optimized dispatchers, and Parallel Folding for flexible parallelism.