Research Briefing — January 29, 2026

Randomized experiments studying how AI assistance affects skill development in programmers learning new libraries. Finds AI use impairs conceptual understanding, code reading, and debugging abilities without significant efficiency gains on average.

Shows reward models inherit significant value biases from their base pretrained LLMs. Demonstrates robust differences along psychological value dimensions (agency vs communion) between Llama and Gemma RMs.

Introduces Self-Distillation Policy Optimization (SDPO) for RLVR that converts rich textual feedback into dense learning signals without external teachers. Treats the model conditioned on feedback as its own teacher.

Proposes SokoBench, a benchmark based on Sokoban puzzles for evaluating long-horizon planning in Large Reasoning Models. Finds consistent degradation beyond 25 moves, suggesting fundamental planning constraints.

Introduces peer prediction methods from mechanism design for LLM evaluation and post-training. Rewards honest and informative answers using mutual prediction between models, enabling evaluation without strong supervision.

Presents quantization-aware distillation (QAD) best practices for recovering accuracy of NVFP4-quantized LLMs and VLMs. Shows effectiveness for models with complex post-training pipelines (SFT+RL+merging) where traditional QAT fails.

Analyzes why multi-agent debate often underperforms majority vote despite higher compute cost. Identifies missing diversity and confidence calibration, proposing lightweight interventions that significantly improve outcomes.

Proposes failure-prefix conditioning to improve learning from saturated problems where standard RLVR stalls. Conditions training on prefixes from rare incorrect trajectories to expose models to failure-prone states.

Proposes new parameterization for MoE transformers enabling hyperparameter transfer when scaling width, depth, number of experts, and expert size. Justified by dynamical mean field theory analysis.