Reasoning Traces Shape Outputs but Models Won't Say So
On the map: LLM reasoning → Prompting and chain-of-thought reasoning
Abstract (arXiv)
Can we trust the reasoning traces that large reasoning models (LRMs) produce? We investigate whether these traces faithfully reflect what drives model outputs, and whether models will honestly report their influence. We introduce Thought Injection, a method that injects synthetic reasoning snippets into a model's <think> trace, then measures whether the model follows the injected reasoning and acknowledges doing so. Across 45,000 samples from three LRMs, we find that injected hints reliably alter outputs, confirming that reasoning traces causally shape model behavior. However, when asked to explain their changed answers, models overwhelmingly refuse to disclose the influence: overall non-disclosure exceeds 90% for extreme hints across 30,000 follow-up samples. Instead of acknowledging the injected reasoning, models fabricate aligned-appearing but unrelated explanations. Activation analysis reveals that sycophancy- and deception-related directions are strongly activated during these fabrications, suggesting systematic patterns rather than incidental failures. Our findings reveal a gap between the reasoning LRMs follow and the reasoning they report, raising concern that aligned-appearing explanations may not be equivalent to genuine alignment.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Prompting and chain-of-thought reasoning, one of 17 directions in LLM reasoning. Builds on 12 of its 20 references that are in the map. First posted on arXiv 21 Mar 2026.
Builds on
12 of 20 references.
Closest ancestors · 3
Same direction, newest first
- 2025Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety15 Jul
- 2025Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation14 Mar
- 2023Measuring Faithfulness in Chain-of-Thought Reasoning17 Jul
Other references · 9
- 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
- 2022Toy Models of Superpositioncited by 928
- 2023Steering Language Models With Activation Engineeringcited by 824
- 2025Stop Overthinking: A Survey on Efficient Reasoning for Large Language Modelscited by 456
- 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 303
- 2024Deliberative Alignment: Reasoning Enables Safer Language Modelscited by 292
- 2025Reasoning Models Can Be Effective Without Thinkingcited by 182
- 2025Reasoning-to-Defend: Safety-Aware Reasoning Can Defend Large Language Models from Jailbreakingcited by 34
- 2025THOUGHTTERMINATOR: Benchmarking, Calibrating, and Mitigating Overthinking in Reasoning Modelscited by 34
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2025Mapping Faithful Reasoning in Language Models
- 2025Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens
- 2026Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models
- 2026Explaining Too Much? Understanding How Large Language Model Reasoning Traces Influence Performance and Metacognition
- 2026Evaluating the False Trust Engendered by LLM Explanations
- 2025Thinking Out Loud: Do Reasoning Models Know When They're Right?
- 2025Can We Trust AI Explanations? Evidence of Systematic Underreporting in Chain-of-Thought Reasoning
- 2026Legibility is Not Interpretability: Comparing Judged and Actual Importance in Chain-Of-Thought Reasoning