Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability
Summary
This paper investigates whether training LLMs for shorter, more efficient chain-of-thought reasoning harms the faithfulness and monitorability of that reasoning. It finds faithfulness generally declines due to reduced model consistency, while monitorability remains relatively robust even with much shorter CoT outputs.
View Cached Full Text
Cached at: 10/05/26, 12:47 PM
Paper page - Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability
Source: https://huggingface.co/papers/2610.03509
Abstract
Chain-of-thought(CoT)reasoningallowshumanstoinspecthowlargelanguagemodelsreachtheiranswers,andoverseemodelbehaviour.Thisreasoningcomesatanincreasedinferencecost,motivatingefficientmethodsthattrainmodelstosolvetasksusingfewertokens.However,acommonconcernisthatsuchtrainingmaycausemodelstoskipimportantreasoningsteps,sotheCoTnolongerfaithfullyreflectsthemodel’sdecision.Itisunclearwhetherorwhenthisoccursinpractice,sincedifferentefficiencymethodsapplylengthpressuretomodels’CoTindistinctways,andfaithfullyexplainingamodel’sdecisiontakesmoretokensonsometasksthanothers.Tounderstandthesedynamics,wefine-tuneavarietyofmodelswiththreemethodsthatapplylengthpressuredifferently,namelyafixedgenerationbudget,aper-examplelengthtarget,andagroup-relativelengthreward.WeevaluatehowefficientreasoningaffectsCoTfaithfulness(i.e.,howwelltheCoTreflectsmodeldecisionsonrelatedinputs)andmonitorability(i.e.,whethertheCoTrevealswheninputinterventionsaltertheoutput).Wefindthatitaffectsfaithfulnessandmonitorabilitydifferently.Faithfulnessfallsinmostsettings,primarilybecausethetrainedmodelsarelessconsistent.Monitorabilityismorerobust,asmodelskeepacknowledgingtheinfluenceontheiranswerevenwhentheCoTismuchshorter.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2610\.03509
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2610.03509 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2610.03509 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2610.03509 in a Space README.md to link it from this page.
Collections including this paper1
Similar Articles
Faithfulness as Information Flow: Evaluating and Training Faithful Chain-of-Thought Reasoning
This paper proposes a framework to evaluate and improve faithfulness of chain-of-thought reasoning by controlling information flow, using entropy-based, KL-divergence, and gradient-based diagnostics, and introduces training interventions (attention masking, gradient masking, adversarial perturbations) that make reasoning more transparent and reduce shortcut reliance.
Learning to Be Monitored: Causal Bounds and Intervention-Aware Post-Training Recipes for Chain-of-Thought Monitorability
Yale researchers establish causal upper and lower bounds on CoT monitorability metrics, showing that the minimal-criterion TPR metric can overestimate intervention-caused behavior detection, and propose three intervention-aware post-training methods (off-policy self-distillation, RL, on-policy self-distillation) that improve LLM reasoning-trace monitorability.
When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models
This paper analyzes failure modes in multi-turn reasoning models by introducing a CoT-Output safety matrix, revealing paradoxes like increased alignment-faking under monitoring cues and context-injection failures where safe internal reasoning is overridden by harmful outputs.
Chain-of-Thought Reasoning in the Wild Is Not Always Faithful
This paper demonstrates that Chain-of-Thought reasoning in large language models is not always faithful, with models producing plausible but incorrect reasoning chains in response to naturally worded prompts, raising concerns for AI safety and transparency.
Reasoning models struggle to control their chains of thought, and that’s good
OpenAI researchers study whether reasoning models can deliberately obscure their chain-of-thought to evade monitoring, finding that current models struggle to control their reasoning even when aware of monitoring. They introduce CoT-Control, an open-source evaluation suite with over 13,000 tasks to measure chain-of-thought controllability in reasoning models.