Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs

arxiv:2609.01117

Original source · Paper JSON

Reasoning and the "physics" of language models · 2026-09-01

Cites (10) · Cited by (0)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4
5. Part 5

Contents

• Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs
• Abstract
• 1 Introduction
• 2 Related Work
• Chain-of-thought and test-time compute.
• Continuous-space reasoning.
• Latent conditioning with a frozen decoder.
• Recurrent and recursive reasoning.
• Iterative reasoning in specialized models.
• 3 Method
• 3.1 Problem Setup
• 3.2 Task-Dedicated Proposer
• Motivation.
• Architecture.
• Cost.
• 3.3 Recurrent Refiner
• States and updates.
• Truncated-gradient unrolling.
• Cost.
• 3.4 Injection and Decoding
• 3.5 Training and Inference
• Two-stage training.
• Refiner unrolling.
• Inference.
• 4 Experiments
• 4.1 Experimental Setup
• Benchmarks.
• Metrics.
• Backbone and baselines.
• What is deliberately not a baseline, and why.
• Reproduction and harness notes.
• Training.
• 4.2 Main Results
• Symbolic reasoning.
• Natural-language reasoning.
• 4.3 Reference Points Outside the Controlled Setting
• Thinking-mode prompting.
• From-scratch symbolic solvers.
• Parameter-efficient prompt adaptation.
• 4.4 Scaling the Recurrent Refiner
• 4.5 Mechanistic Ablations
• Factorial decomposition.
• Depth versus capacity.
• Refinement is incremental.
• 4.6 Where Does the Computation Happen?
• 5 Conclusion
• Limitations
• Per-task modules, and what that costs the claim.
• Generalization is within-task, not across tasks.
• Scope of evaluation.
• Method caveats.
• Appendix A Implementation and Training Details
• Architecture.
• Optimization.
• Datasets.
• What the supervision consists of.
• Inference.
• Appendix B Reproduction and Harness Notes
• Baseline reproduction.
• Thinking versus non-thinking mode.
• Adapting TRM to Countdown-4.
• Appendix C Parameter-Efficient Baselines and Data Efficiency
• Do soft prompts alone explain the gain?
• Is the gain bought with 100k synthetic examples?
• Appendix D Refiner Design Ablations
• Appendix E Scaling and Depth
• Test-time depth extrapolation.
• Non-recurrent control.
• Latent count and width.
• Why K = 32 K{=}32 in training but K infer = 4 K_{\text{infer}}{=}4 at test?
• Appendix F Training Protocol Ablations
• Appendix G Mechanistic Analyses
• Refinement trajectory.
• Latent convergence.
• What the latents encode.
• Linear probe on the refined latents.
• Why generic latents collapse.
• Appendix H Generality Across Decoders
• Appendix I Compute and Efficiency
• Training cost.
• Appendix J Potential Risks
• Appendix K Licenses, Parameters, and Intended Use of Artifacts
• K.1 Existing Artifacts
• Pretrained models.
• Baseline code.
• Evaluation datasets.
• Generated data.
• Software packages and parameter settings.
• K.2 Artifacts We Create
• Appendix L Use of AI Assistants
• Instructions for reporting errors