SuCo introduces Minimal Sufficient CoT (MSC) as the shortest reasoning prefix adequate for correct answers. It employs a two-stage training framework—MSC-Aligned Fine-Tuning and Sufficiency-Aware Policy Optimization—to reduce reasoning length while maintaining or improving accuracy across math, code, and science tasks.
Dynamic Rollout Editing (DRE) addresses overthinking in RL-trained reasoning models by modifying successful trajectories post-answer emergence. DRE preserves the correct reasoning prefix while editing unnecessary continuation, weakening the credit assigned to redundant thinking without penalizing valid reasoning. Experiments across diverse tasks demonstrate its effectiveness in reducing overthinking.
DeepRubric introduces a data construction framework that builds query-rubric pairs by first defining verifiable evaluation targets through an evidence tree. It generates 9K supervision examples and trains a 8B model with GRPO, achieving performance comparable to state-of-the-art models using 13x fewer RL GPU-hours.
ExpRL introduces a novel mid-training approach for LLMs using human-written question-answer data as reward scaffolds. Instead of imitating reference solutions, it constructs problem-specific grading rubrics to reward intermediate reasoning steps, enabling better initialization for sparse-reward RL and outperforming SFT, sparse-reward GRPO, and self-distillation on math reasoning tasks.
Hierarchical Advantage-Weighted Behavior Cloning (HABC) enhances online RL fine-tuning of vision-language agents by using separate critic heads for viability and efficiency. It combines their outputs via a state-adaptive gate and applies per-transition weights, while intervention-aware credit assignment prevents supervision leakage. In real-robot experiments, HABC boosts success rates to 92%, 88%, and 38% on three bimanual tasks, surpassing SFT baselines of 36%, 44%, and 12%.
The Geometric Action Model (GAM) enables robot policies to reason about 3D physical interactions by repurposing a pretrained geometric foundation model. GAM splits the GFM to serve as both an observation encoder and a causal future predictor, then routes predicted future geometry and actions through the same backbone, achieving accurate, robust, and efficient manipulation performance in simulation and real-robot benchmarks.
The podcast reviews the evolution of post-training recipes in large language models, from InstructGPT to 2026 frontier models. It highlights Multi-Teacher On-Policy Distillation (MOPD) as the dominant pattern, where domain-specialist models are trained and then distilled into a general student model via on-policy distillation, scaling to over 10 teachers in models like DeepSeek V4 and Nemotron 3 Ultra.
A training-free diagnostic, contrastive-difference CKA (CKA_Delta), identifies concept-specific structural alignment across language model architectures. It detects geometric convergence and functional transfer across six concept domains, including non-instructional tasks, with significant discrimination where standard CKA fails. Results suggest universality may strengthen with model scale, though further validation is needed.
A study identifies extrinsic (crucial tokens) and intrinsic (cognitive behaviors) properties that enhance code interpreter reasoning in large language models. Stronger reasoning models show higher prevalence of verification, backtracking, and backward chaining, with these properties improving performance during inference and training, reducing overthinking and boosting token efficiency.
MA-SBI introduces a calibration-free simulation-based inference framework that uses side-channel text, like regime labels or instructions, to correct for simulator misspecification. It employs a learned corrector to apply observation-space shifts before posterior inference, without needing ground-truth parameter pairs or retraining. On hide-the-calibration benchmarks, MA-SBI matches the oracle posterior with text alone, outperforming RoPE under limited data, and shows robustness on real-world epidemiological and cognitive-science datasets.
This paper introduces a unified causal-origin taxonomy that categorizes distributional shifts in reinforcement learning into internal, agent-driven, and external, environment-driven sources. It unifies ID/OOD generalization and non-stationary settings by framing shifts as structured changes in the agent-environment interaction process, using a POMDP decomposition and a shifted-time boundary perspective.
A quality cliff at 6.25 Hz in neural audio codecs is caused by insufficient training token exposure due to fixed clip duration. Correcting this training configuration enables smooth WER degradation down to 3.1 Hz and 1.6 Hz, indicating low frame rate efficiency is more achievable than previously thought.
PACT combines a reactive RL policy with a 2B-parameter Small Language Model to generate and validate action plans. The SLM plan is executed directly if verified as safe, feasible, and complete, bypassing the RL policy. PACT outperforms baselines on three increasingly difficult FrozenLake environments.
Hyperball is a simple optimizer wrapper that sets fixed Frobenius norms for weight matrices and their updates. It improves training speed and learning rate transfer in large models, achieving 20--30% token equivalent speedup over weight decay baselines on up to 1.2B parameter models.
Factorized Neural Operators (FaNO) decompose spectral representations into equivariant dynamic and invariant persistent responses. This factorized structure enables better interpretability, generalization, and consistent predictions across scales, domains, and physical regimes.
We propose a new functional gradient descent algorithm that adapts its representation during optimization. The method achieves convergence to a stationary point under smooth losses and to a global minimizer under smoothness and a Polyak-Lojasiewicz condition, despite using finite-dimensional approximations. It outperforms both fixed-approximation FGD and neural network baselines in regression, PDE solving, and computer vision tasks.
This paper proposes a unified causal-origin taxonomy for distributional shifts in reinforcement learning, linking ID/OOD generalization to non-stationary settings. It decomposes the agent-environment interaction using a POMDP framework, identifying internal, agent-driven, and external, environment-driven shifts, with explicit, implicit, and hybrid types defined by the shifted-time boundary. The work introduces an evaluation framework to measure shift impact through performance degradation and recovery metrics, enabling systematic analysis of RL robustness.
A study identifies extrinsic (crucial tokens) and intrinsic (cognitive behaviors) properties that enhance code interpreter reasoning in large language models. Stronger reasoning models show higher prevalence of verification, backtracking, and backward chaining, with these properties improving performance during inference and training, reducing overthinking and boosting token efficiency.
PACT combines a reactive RL policy with a 2B-parameter Small Language Model to generate and validate action plans. The SLM plan is executed directly if verified in simulation, bypassing the RL policy without retraining. PACT outperforms baselines on three increasingly difficult FrozenLake environments.
A measurement study finds that 26 semantic post-hoc operators do not improve held-out accuracy over Best-of-N in frozen small code models. While some operators reduce compute usage or recover correct programs, none outperform BoN in accuracy, due to systemic limitations like coverage walls and consensus traps. An expression-layer recovery (M1) improves performance on HumanEval+ by 12 tasks, with no harm or leakage, and shows consistent results across model cells.