Length Penalties Make Chain-of-Thought Less Monitorable
On the map: LLM reasoning → Test-time compute and efficient thinking
Abstract (arXiv)
Recent work trains reasoning models with length penalties to curb overthinking and cut inference cost. We show that these penalties make the chain of thought less monitorable. A length-compressed model still lets misleading hints steer its answers, but it less often verbalizes their influence. We train Qwen3-4B and Qwen3-14B with reinforcement learning under length penalties targeting 60% down to 30% of baseline chain-of-thought length, then evaluate them with nine types of biasing hints on held-out MMLU-Pro-R and four transfer benchmarks. A chain is faithful when an LLM monitor can tell from it that the hint influenced the answer. At the 30% target, accuracy stays near baseline and wrong-answer hints switch answers as often as before. Yet faithfulness drops on every evaluation set for both models, by 39% for Qwen3-14B and 35% for Qwen3-4B on MMLU-Pro-R. A control trained with the same correctness and format rewards but no length penalty leaves faithfulness intact or raises it. Shortening alone does not explain the drop. Compressed chains mention the hint 7 to 35 percentage points less often than the uncompressed model's chains shortened to the same length by random sentence deletion, across both model sizes and all five evaluation sets. Length penalties therefore trade monitorability for inference cost by removing the evidence monitors depend on.
Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 40 of its 50 references that are in the map. First posted on arXiv 8 Jul 2026.
Builds on
40 of 50 references.
Closest ancestors · 8
Same direction, newest first
- 2025DeepCompress: A Dual Reward Strategy for Dynamically Exploring and Compressing Reasoning Chains31 Oct
- 2025NVIDIA Nemotron Nano 2: An Accurate and Efficient Hybrid Mamba-Transformer Reasoning Model20 Aug
- 2025Just Enough Thinking: Efficient Reasoning with Adaptive Length Penalties Reinforcement Learning5 Jun
- 2025ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning30 Apr
- 2025ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning2 Apr
- 2025L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning6 Mar
- 2025s1: Simple test-time scaling31 Jan
- 2024Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs30 Dec
Other references · 32
- 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,302
- 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
- 2025Qwen3 Technical Reportcited by 8,195
- 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
- 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
- 2022Large Language Models are Zero-Shot Reasonerscited by 4,323
- 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,678
- 2024HybridFlow: A Flexible and Efficient RLHF Frameworkcited by 1,980
- 2024MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmarkcited by 1,758
- 2020What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Examscited by 1,324
- 2025Understanding R1-Zero-Like Training: A Critical Perspectivecited by 1,276
- 2023Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Promptingcited by 1,081
20 more
- 2023Measuring Faithfulness in Chain-of-Thought Reasoningcited by 668
- 2019MathQA: Towards Interpretable Math Word Problem Solving with Operation-Based Formalismscited by 653
- 2025Reasoning Models Don't Always Say What They Thinkcited by 388
- 2025Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscationcited by 338
- 2023Faithful Chain-of-Thought Reasoningcited by 285
- 2025Training Language Models to Reason Efficientlycited by 242
- 2025Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safetycited by 222
- 2020ReClor: A Reading Comprehension Dataset Requiring Logical Reasoningcited by 211
- 2025Chain-of-Thought Reasoning In The Wild Is Not Always Faithfulcited by 166
- 2025When Chain of Thought is Necessary, Language Models Struggle to Evade Monitorscited by 81
- 2025Are DeepSeek R1 And Other Reasoning Models More Faithful?cited by 69
- 2025CoT Red-Handed: Stress Testing Chain-of-Thought Monitoringcited by 49
- 2025Monitoring Monitorabilitycited by 47
- 2024MMLU-CF: A Contamination-free Multi-task Language Understanding Benchmarkcited by 41
- 2025Stable Reinforcement Learning for Efficient Reasoningcited by 27
- 2026Reasoning Models Struggle to Control their Chains of Thoughtcited by 22
- 2025Reasoning Under Pressure: How do Training Incentives Influence Chain-of-Thought Monitorability?cited by 7
- 2026MonitorBench: A Comprehensive Benchmark for Chain-of-Thought Monitorability in Large Language Modelscited by 7
- 2026Aligned, Orthogonal or In-conflict: When can we safely optimize Chain-of-Thought?cited by 3
- 2026Chain-of-thought obfuscation learned from output supervision can generalise to unseen taskscited by 2
Built on by
Citing papers per year: 2026: 1.
1 paper.
Similar, not cited
Close in wording, no citation either way
- 2026Stepwise Penalization for Length-Efficient Chain-of-Thought Reasoning
- 2026To Think or Not to Think: Allocating Reasoning Where It Helps
- 2026On the Optimal Reasoning Length for RL-Trained Language Models
- 2025When More is Less: Understanding Chain-of-Thought Length in LLMs
- 2025How Well do LLMs Compress Their Own Chain-of-Thought? A Token Complexity Approach
- 2025Thinking Fast and Right: Balancing Accuracy and Reasoning Length with Adaptive Rewards
- 2026Quantized Reasoning Models Think They Need to Think Longer, but They Do Not
- 2026Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning