Diffusion Models Meet Reasoning: Learning to Schedule, Search, and Synthesize

This week brings advances in how we train and optimize AI systems across reasoning, robotics, and multimodal domains. We‘re seeing a shift toward learning when and how to use compute primitives, from scheduling token generation in diffusion models to adaptively invoking code in vision-language systems. Simultaneously, researchers are cracking long-standing problems in model composition, safety alignment for robotic systems, and the surprising ineffectiveness of naive prompt optimization in multi-agent setups. The papers also reveal emerging tensions: training diversity improves safety but not semantic reasoning; longer context training helps but requires careful curriculum design; and compression techniques benefit substantially from second-order optimization principles.

Here‘s what caught our attention:

• SPIRAL formulates multi-trace reasoning as a set RL problem, achieving 11× higher scaling efficiency by training models to generate diverse parallel reasoning paths whose usefulness is determined collectively rather than individually, a fundamental rethinking of inference compute.

• Scheduling Thoughts derives an information-theoretic upper bound on diffusion language model performance and uses it to train an optimal unmasking policy via GRPO, improving Sudoku accuracy from 82% to 91.8% with a frozen base model.

• dVLA-RL solves the intractable problem of computing action probabilities in discrete diffusion policies by reformulating to trajectory-level probabilities, enabling RL optimization of vision-language-action models with 30.6pp gains on bimanual manipulation.

• VeriEvol decouples prompt evolution from answer verification as independent pipeline axes, treating verification as a dataset property rather than a training concern, scaling SFT data 25× while maintaining reliability through multi-source falsification.

• AIR introduces group-constrained RL rewards that decouple tool invocation from accuracy rewards, preventing agentic training collapse while achieving 95%+ code execution reliability across mathematical reasoning benchmarks.

• Randomized YaRN improves length generalization through randomized position sampling with curriculum-based training, achieving 90% accuracy on out-of-distribution 128K contexts, revealing the critical importance of curriculum rather than static extrapolation.

• SVD-Surgeon brings Optimal Brain Surgeon to the singular-value basis with closed-form updates, substantially improving low-rank compression trade-offs without retraining (e.g., 70% compression perplexity: 944→46 on OPT-6.7B).

• LIBERO-Safety establishes systematic evaluation of physical and semantic safety in VLAs through parametric scenario generation, revealing that the “generalization-safety tension” prevents existing models from scaling collision-free trajectory synthesis.

Let‘s get into it 👇

Bi-Weekly AI Research Roundup

Latest research summaries in ML, Robotics, CV, NLP and AI

Contents

1. SPIRAL: Learning to Search and Aggregate

2. Teaching LLMs String Matching, Backtracking, and Error Recovery to Deduce Bases and Truth Tables for the Combinatorially Exploding Bit Manipulation Puzzles

3. Controllable Accent Normalization via Discrete Diffusion

4. VeriEvol: Scaling Multimodal Mathematical Reasoning via Verifiable Evol-Instruct

5. AIR: Adaptive Interleaved Reasoning with Code in MLLMs

6. PISCES: Annotation-free Text-to-Video Post-Training via Optimal Transport-Aligned Rewards

7. Scaling Linear Mode Connectivity and Merging to Billion Parameter Pretrained Transformers

8. MAS-PromptBench: When Does Prompt Optimization Improve Multi-Agent LLM Systems?

9. Scheduling Thoughts: Learning the Order of Thought in Diffusion Language Models

10. Randomized YaRN Improves Length Generalization for Long-Context Reasoning

11. SVD-Surgeon: Optimal Singular-Value Surgery for Large Language Model Compression

12. LangMAP: A Language-Adaptive Approach to Tokenization

13. dVLA-RL: Reinforcement Learning over Denoising Trajectories for Discrete Diffusion Vision-Language-Action Models

14. LIBERO-Safety: A Comprehensive Benchmark for Physical and Semantic Safety in Vision-Language-Action Models

15. Learning to See While Learning to Act: Diffusion Models for Active Perception in Robot Imitation

SPIRAL: Learning to Search and Aggregate

Authors: Jubayer Ibn Hamid, Ifdita Hasan Orney, Michael Y. Li, Omar Shaikh, Yoonho Lee, Dorsa Sadigh, Chelsea Finn, Noah Goodman