mlsanityData to 14 Oct 2026

Base Models Can Reason By Taking a Cue From Training Data

arXiv 2610.06851 ↗First posted 5 Oct 2026cs.LG cs.AI cs.CLcited by 0

On the map: LLM reasoning → Prompting and chain-of-thought reasoning also fits: Reinforcement learning for reasoning

Abstract (arXiv)

In this paper, we study how training data creates associations between the tokens at the start of a base model's response and the reasoning behavior that follows. First, we demonstrate that fixing particular starting token cues makes a base model's performance competitive with that of its reinforcement learning (RL)-trained counterparts on math and coding. For instance, the cue ".nnOkay" raises Olmo-3-7B's MATH-500 pass@1 accuracy from 42% to 78%, while "Alright," raises Qwen3-14B's from 72% to 87%. Second, RL makes these cues more likely, while fixing them recovers much of its performance gain over the base model. Third, we trace the reasoning effects of token cues to the training data. We perform causal data interventions to turn an arbitrary word, such as "chicken", into an effective reasoning cue, or remove an existing cue's effect. A similar edit makes the prompt instruction "Think duck duck goose" as effective as "Think step by step" at eliciting reasoning. We also find that the hidden state representations induced by different cues correlate with different document types from the training set. Finally, we extend our study of token cues with a case study in language model safety, finding that different cues elicit distinct refusal and compliance behaviors that correspond to different types of training data.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 13 of its 67 references that are in the map. First posted on arXiv 5 Oct 2026.

Builds on

13 of 67 references.

References · 13

  1. 2014Adam: A Method for Stochastic Optimizationcited by 37,240
  2. 2024The Llama 3 Herd of Modelscited by 14,715
  3. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  4. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  5. 2025Qwen3 Technical Reportcited by 8,820
  6. 2021Evaluating Large Language Models Trained on Codecited by 8,320
  7. 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 3,104
  8. 2024Qwen2.5-Math Technical Report: Toward Mathematical Expert Model via Self-Improvementcited by 1,090
  9. 2019CTRL: A Conditional Transformer Language Model for Controllable Generationcited by 762
  10. 2023Studying Large Language Model Generalization with Influence Functionscited by 326
  11. 2025Olmo 3cited by 129
  12. 2025OctoThinker: Mid-training Incentivizes Reinforcement Learning Scalingcited by 113
1 more
  1. 2026The Truth Lies Somewhere in the Middle (of the Generated Tokens)cited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models
  2. 2025Internal states before wait modulate reasoning patterns
  3. 2025Base Models Know How to Reason, Thinking Models Learn When
  4. 2026Extremely Sparse Supervision Incentivizes Reasoning Ability
  5. 2025Thinking Out Loud: Do Reasoning Models Know When They're Right?
  6. 2026Probing the Origins of Reasoning Performance: Representational Quality for Mathematical Problem-Solving in RL vs. SFT Fine-Tuned Models
  7. 2026How RL Reshapes LLM Reasoning: Transferability, Coverage, and Scaling Laws
  8. 2025Scaling Reasoning, Losing Control: Evaluating Instruction Following in Large Reasoning Models