mlsanityData to 14 Oct 2026

Transformers Stop Thinking Too Early, and a Tiny LoRA Fixes It

arXiv 2609.36585 ↗First posted 29 Sep 2026cs.AI cs.CL cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Pretrained transformers use little of their depth to follow references in context. Thirteen base models reliably follow only 1.4-3.6 lines, and extra pretrained loops add little. A task-trained rank-8 LoRA at one early layer extends this computation with all model weights frozen. Qwen3-8B improves from 15.5% to 99% exact accuracy on 24-line chains; a longer-trained LoRA reaches 50 lines. Ouro-1.4B reaches 60 lines after four loops and at least 160 after eight. The LoRA starts a relay: program lines pass on their chain identity through a short range of middle layers. Frozen heads read progressively further up the chain, and removing parent-line attention stops the relay. A frozen-model measurement locates the last useful intervention layer within tolerance in three of four held-out models. Task-specific LoRAs also improve MuSiQue. Default answers therefore understate the computation accessible through a tiny edit. Code and an interactive demo are available at https://lunamos.github.io/stop-thinking-too-early/

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 60 of its 63 references that are in the map. First posted on arXiv 29 Sep 2026.

Builds on

60 of 63 references.

References · 60

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2025Qwen3 Technical Reportcited by 8,820
  3. 2021The Power of Scale for Parameter-Efficient Prompt Tuningcited by 3,795
  4. 2016Pointer Sentinel Mixture Modelscited by 3,096
  5. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  6. 2025Gemma 3 Technical Reportcited by 1,824
  7. 2024RULER: What's the Real Context Size of Your Long-Context Language Models?cited by 1,346
  8. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  9. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  10. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  11. 2022Measuring and Narrowing the Compositionality Gap in Language Modelscited by 1,075
  12. 2022In-context Learning and Induction Headscited by 964
38 more
  1. 2023Steering Language Models With Activation Engineeringcited by 922
  2. 2021MuSiQue: Multihop Questions via Single-hop Question Compositioncited by 916
  3. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 797
  4. 2020Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuningcited by 671
  5. 2018Universal Transformerscited by 647
  6. 2023Dissecting Recall of Factual Associations in Auto-Regressive Language Modelscited by 528
  7. 2025Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approachcited by 415
  8. 2022Transformers Learn Shortcuts to Automatacited by 268
  9. 2025Reasoning with Latent Thoughts: On the Power of Looped Transformerscited by 240
  10. 2023Looped Transformers as Programmable Computerscited by 237
  11. 2024ReFT: Representation Finetuning for Language Modelscited by 197
  12. 2024Let's Think Dot by Dot: Hidden Computation in Transformer Language Modelscited by 191
  13. 2024Do Large Language Models Latently Perform Multi-Hop Reasoning?cited by 169
  14. 2025Scaling Latent Reasoning via Looped Language Modelscited by 165
  15. 2023Looped Transformers are Better at Learning Learning Algorithmscited by 149
  16. 2024Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Trackingcited by 149
  17. 2025Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computationcited by 133
  18. 2021Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networkscited by 117
  19. 2023How do Language Models Bind Entities in Context?cited by 108
  20. 2022Unveiling Transformers with LEGO: a synthetic reasoning taskcited by 106
  21. 2024Looped Transformers for Length Generalizationcited by 99
  22. 2024Transformers, parallel computation, and logarithmic depthcited by 79
  23. 2024Hopping Too Late: Exploring the Limitations of Large Language Models on Multi-Hop Queriescited by 78
  24. 2024LoFiT: Localized Fine-tuning on LLM Representationscited by 57
  25. 2025Do Language Models Use Their Depth Efficiently?cited by 50
  26. 2024A Mechanistic Analysis of a Transformer Trained on a Symbolic Multi-Step Reasoning Taskcited by 49
  27. 2025Teaching Pretrained Language Models to Think Deeper with Retrofitted Recurrencecited by 47
  28. 2023Entity Tracking in Language Modelscited by 44
  29. 2025Language Models use Lookbacks to Track Beliefscited by 39
  30. 2026Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformerscited by 37
  31. 2022End-to-end Algorithm Synthesis with Recurrent Networks: Logical Extrapolation Without Overthinkingcited by 33
  32. 2026A Mechanistic Analysis of Looped Reasoning Language Modelscited by 33
  33. 2025Latent Chain-of-Thought? Decoding the Depth-Recurrent Transformercited by 27
  34. 2025Encode, Think, Decode: Scaling test-time reasoning with recursive latent thoughtscited by 24
  35. 2025Back Attention: Understanding and Enhancing Multi-Hop Reasoning in Large Language Modelscited by 23
  36. 2025How Do Transformers Learn Variable Binding in Symbolic Programs?cited by 21
  37. 2025How Do LLMs Perform Two-Hop Reasoning in Context?cited by 20
  38. 2024Transformers Struggle to Learn to Searchcited by 19

50 most cited of 60 All 60 (JSON)

Built on by

Not cited yet.