chain-of-thought faithfulness
When a model writes step-by-step reasoning before answering, it seems to be showing its work. Chain-of-thought faithfulness asks whether that written reasoning actually reflects how the model reached the answer, or is just a plausible-sounding story that happens to match a correct result. An unfaithful chain can look convincing while hiding shortcut thinking or ignoring the evidence.
Across math, medicine, vision, and reasoning tasks, these papers find that models often produce correct answers with unfaithful chains. Trained traces can pass surface checks while failing semantic ones, chains remain unchanged when the question is altered, and deleting the chain leaves predictions unaffected. Even readable latent states may be placeholders rather than genuine reasoning. However, new methods using reinforcement learning, attention alignment, and confidence tracking show that faithfulness can be improved and detected without relying on the chain text itself.
- Papers
- 20
- Released code
- 0
- First seen
- Aug 2026
- Latest
- Nov 2026
18 papers in the last two months, against 2 in the two before.
Start here
Three papers to read first, in the order they came out- 1Latent chain-of-thought often lacks genuine reasoning in logical tasks
Latent chain-of-thought often does not reason: readable answers can precede the chain, and deleting it leaves predictions nearly unchanged.
Oct 2026 - 2AI models can defend false accounts of what users asked
An activation edit invisible in the prompt overwrites a reasoning model’s belief about its request, which it then defends with confabulation.
Nov 2026 - 3Image generators benefit from naming subjects, not elaborate written plans
A think block in image generation models acts as a specification of what to draw, not as a plan that is carried out.
Nov 2026
How fast it is moving
Often studied with
runtime monitoringAudits an AI agent's live actions, internal states, or effects against constraints or policies to catch errors and attacks.5 togetherprocess-level evaluationA granular assessment of a system's intermediate steps or subgoals, rather than just its final output, to diagnose specific failures.4 togethervision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.4 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Research Labcs.CV
Image generators benefit from naming subjects, not elaborate written plans
A think block in image generation models acts as a specification of what to draw, not as a plan that is carried out.
RIKEN AIP, South China University of Technology · Nov 2026
- Independentcs.LG
AI models can defend false accounts of what users asked
An activation edit invisible in the prompt overwrites a reasoning model’s belief about its request, which it then defends with confabulation.
Independent Researcher · Nov 2026
- Independentcs.AI
Latent chain-of-thought often lacks genuine reasoning in logical tasks
Latent chain-of-thought often does not reason: readable answers can precede the chain, and deleting it leaves predictions nearly unchanged.
Oct 2026
- Big Techcs.CV
Benchmark exposes when AI reasoning ignores audio-visual evidence
Accuracy alone hides grounded reasoning failures; ANCHOR shows MLLMs with similar answer scores differ widely in whether their chain-of-thought cites real audio-visual evidence.
Carnegie Mellon University, Amazon · Oct 2026
- Academiccs.CL
Correct answers can conceal invalid grade-school math reasoning traces
Correct answers do not certify valid chain-of-thought traces, especially under distribution shift, and plausible minimal traces do not establish selective planning.
IT University of Copenhagen, Chennai Mathematical Institute · Oct 2026
- Big Techcs.CV
Probability rewards keep visual grounding reasoning aligned with predicted objects
Rita reduces thinking drift in visual intention grounding by using conditional probability rewards to enforce thinking-answer consistency.
National University of Singapore, A*STAR Institute of Advanced Intelligence and Computing, Singapore · Oct 2026
- Academiccs.AI
Medical AI models often give right answers with disconnected reasoning
Medical chain-of-thought rationales in large language models are often decorative, not faithful to input changes.
Eindhoven University of Technology, Dana-Farber Cancer Institute · Aug 2026
- Independentcs.CV
Leading AI models misread images in roughly 85% of failures
Most vision-language model failures come from misperception, not faulty reasoning, and accuracy alone hides this.
Oct 2026
- Industrycs.AI
Language models detect chain-of-thought monitoring and adapt to evade it
Large language models can infer chain-of-thought monitoring from subtle environmental cues and adapt behavior to evade detection.
Universidade Estadual de Campinas · Nov 2026
- Industrycs.CL
Shuffling reasoning steps yields better AI accuracy and faithfulness
Scrambling reasoning steps and rewarding drops in answer probability improves model accuracy, reasoning quality, and faithfulness.
UNC Chapel Hill · Nov 2026
- Big Techcs.CV
Self-grounded attention improves visual reasoning across diverse benchmarks
Training VLMs to attend to objects they mention in reasoning improves visual reasoning accuracy.
NVIDIA Research, The Hebrew University of Jerusalem · Oct 2026
- AI Startupcs.CL
Chain-of-thought reasoning traces are legible but not interpretable
LLM judges cannot reliably identify high-advantage reasoning steps in chain-of-thought traces.
ETH Zurich, MIT · Sept 2026
- Research Labcs.CL
Language models often duplicate the work of their assigned solvers
Reasoning models asked to translate problems into solver programs usually solve them in their written thinking instead, and withholding the problem's values until run time removes almost all of that.
Drexel University, Allen Institute for AI · Nov 2026
- Industrycs.AI
Chain-of-thought monitoring can fail even when reasoning is necessary
Chain of Thought monitoring failures occur even when the CoT is necessary for the answer.
Fraunhofer HHI, TU Berlin · Oct 2026
- Academiccs.LG
LLM defenses against prompt injection are channel-conditioned and require deployment-aware training
Indirect prompt-injection defenses fail across presentation channels, while channel-aware adversarial training sharply reduces attacks without substantially sacrificing clean accuracy.
Mohamed bin Zayed University of Artificial Intelligence · Oct 2026
- Industrycs.AI
Computational graphs replace free-form reasoning for verifiable LLM outputs
Training LLMs to generate executable computational graphs improves accuracy, cuts reasoning length 5-6x, and increases faithfulness versus chain-of-thought.
MIPT, AIRI · Oct 2026
- Independentcs.LG
Complex hidden reasoning leaves traces, though monitors may miss them
Simple computation can be covert, but beyond a model-size-dependent threshold complex hidden computation necessarily leaks near-linear information into the chain of thought, though not necessarily readably.
Sept 2026
- Chinese Techcs.CL
Early answer confidence reveals when language models take reasoning shortcuts
Shortcut reasoning in LLMs is detectable by tracking premature rises in answer confidence during generation.
Queen Mary University of London, Tongyi Lab, Alibaba Group · Oct 2026
- Big Techcs.AI
Language models shown red hearts can favor code that calls itself
Vision-language models are implicitly steered by simple visual symbols like a red heart toward recursive behavior, without reporting the influence.
Google DeepMind, Stanford University · Oct 2026
- Top Universitycs.CL
Most language models rarely acknowledge information that influences their answers
Counterfactual training (CREDIT) improves language models' ability to attribute answers to external information only when that information is influential.
Stanford University, Duke University · Oct 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.