REVIEW 32 cited by
Reasoning Models Know When They're Right: Probing Hidden States for Self-Verification
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Reasoning Models Know When They're Right: Probing Hidden States for Self-Verification
read the original abstract
Reasoning models have achieved remarkable performance on tasks like math and logical reasoning thanks to their ability to search during reasoning. However, they still suffer from overthinking, often performing unnecessary reasoning steps even after reaching the correct answer. This raises the question: can models evaluate the correctness of their intermediate answers during reasoning? In this work, we study whether reasoning models encode information about answer correctness through probing the model's hidden states. The resulting probe can verify intermediate answers with high accuracy and produces highly calibrated scores. Additionally, we find models' hidden states encode correctness of future answers, enabling early prediction of the correctness before the intermediate answer is fully formulated. We then use the probe as a verifier to decide whether to exit reasoning at intermediate answers during inference, reducing the number of inference tokens by 24\% without compromising performance. These findings confirm that reasoning models do encode a notion of correctness yet fail to exploit it, revealing substantial untapped potential to enhance their efficiency.
Forward citations
Cited by 32 Pith papers
-
LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling
AutoTTS discovers width-depth test-time scaling controllers through agentic search in a pre-collected trajectory environment, yielding better accuracy-cost tradeoffs than hand-designed baselines on math reasoning task...
-
Plans Don't Persist: Why Context Management Is Load Bearing for LLM Agents
Replay pairing shows LLM agents do not persist plans in hidden states but rely on plans remaining in context, with rapid signal decay and task performance drops when plans are evicted.
-
When Agents Commit Too Soon: Diagnosing Premature Commitment in LLM Agents
Hidden-state convergence at step 4 predicts behavioral consistency in LLM agents on QA tasks (r=-0.35 to -0.83), enabling AUROC 0.97 detection of inconsistent trajectories but not improving accuracy on harder benchmarks.
-
Grounded or Guessing? LVLM Confidence Estimation via Blind-Image Contrastive Ranking
BICR trains a lightweight probe on contrastive hidden states from real versus blind images to detect visual grounding in LVLM predictions, outperforming baselines on calibration and discrimination with fewer parameters.
-
Grounded or Guessing? LVLM Confidence Estimation via Blind-Image Contrastive Ranking
BICR uses blind-image contrastive ranking on frozen LVLM hidden states to train a lightweight probe that penalizes confidence on blacked-out inputs, yielding top calibration and discrimination across five models and m...
-
LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling
AutoTTS discovers superior test-time scaling strategies for LLMs via cheap controller synthesis in a pre-collected trajectory environment, outperforming manual baselines on math benchmarks with low discovery cost.
-
Do Prompt-Elicited Trajectories Reflect Training-Time Reward Hacking? A Systematic Study on Monitoring Training-Time Reward Hacking in Code Generation
Monitors trained on prompt-elicited reward-hacking trajectories fail to generalize to hacking behaviors that arise naturally during RL training of code models, whereas trajectories curated by Trace-and-Amplify transfe...
-
From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning
SFT supplies entangled compositional traces of atomic skills and routing modules; RL identifies those modules and enables recombination on novel compositions outside the SFT support.
-
Multi-Head Latent Control: A Unified Interface for LLM Agent Decision Making
Hidden-state traces of frozen LLMs/VLMs can be read by lightweight trained heads to predict when to defer, clarify, call tools, or abstain, cutting routed inference cost 27–90%.
-
LLMs as a Jury: Cross-Model Consensus Can Outperform Process Reward Models for LLM Reasoning
Cross-model answer agreement is a free Best-of-N verifier that matches trained process reward models and is predicted by a parameter-free law from three panel statistics.
-
LLMs as a Jury: Cross-Model Consensus Can Outperform Process Reward Models for LLM Reasoning
A panel of independent LLMs whose answers agree is a stronger, training-free verifier than self-consistency or process reward models, with a closed-form law predicting its accuracy and its shared-error ceiling.
-
Does the Same Token Mean the Same State? MoE Routing as Signal for Reasoning Control
MoE routing states at boundary and delimiter anchors form basins that align with final answers, enabling RAD, a string-free multi-rollout selector that matches majority voting on math and code tasks.
-
From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning
Introduces a hierarchical latent selection model showing SFT supplies raw module materials in compound traces while RL decomposes them to identify atomic modules and enable recombination for new reasoning configurations.
-
Dynamic Rollout Editing for Reducing Overthinking in RL-Trained Reasoning Models
Dynamic Rollout Editing reduces overthinking in RL-trained LLMs by editing post-answer continuations in successful rollouts and preferring the edited versions within GRPO groups.
-
When is Your LLM Steerable?
Early hidden state features from the first few tokens allow a GBDT classifier to predict activation steering success, under-steering, or over-steering with 0.7 macro-F1 on unseen concepts.
-
Small RL Controller, Large Language Model: RL-Guided Adaptive Sampling for Test-Time Scaling
RL-trained lightweight controller using answer statistics improves trade-offs among correctness, latency, and total samples in adaptive sampling for LLM test-time scaling.
-
Hidden Error Awareness in Chain-of-Thought Reasoning: The Signal Is Diagnostic, Not Causal
LLMs detect CoT reasoning errors in hidden states with 0.95 AUROC but cannot use this awareness to correct them via steering, patching, or self-correction, indicating the signal is diagnostic not causal.
-
Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight
Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to cut up to 50% wasted reasoning tokens and recover safe actions from 80% of unsafe traces, more than doubling success rat...
-
Behavior Cue Reasoning: Monitorable Reasoning Improves Efficiency and Safety through Oversight
Behavior Cue Reasoning trains LLMs to emit special tokens before behaviors, enabling monitors to prune up to 50% of wasted tokens and recover safe actions from 80% of unsafe traces, more than doubling success rates wi...
-
Spatiotemporal Hidden-State Dynamics as a Signature of Internal Reasoning in Large Language Models
Large reasoning models show measurable hidden-state dynamics that a new statistic can use to distinguish correct reasoning trajectories without labels.
-
Do Prompt-Elicited Trajectories Reflect Training-Time Reward Hacking? A Systematic Study on Monitoring Training-Time Reward Hacking in Code Generation
Prompt-elicited hacking trajectories do not reflect training-time reward hacking in code generation; monitors trained on Trace-and-Amplify data generalize better to unseen hacking types.
-
Do Prompt-Elicited Trajectories Reflect Training-Time Reward Hacking? A Systematic Study on Monitoring Training-Time Reward Hacking in Code Generation
Synthetic reward hacking data does not capture natural hacking behaviors in code generation RL, causing monitors trained on it to generalize poorly compared to those trained on in-the-wild trajectories.
-
Emergent Manifold Separability during Reasoning in Large Language Models
Reasoning in LLMs produces a transient geometric pulse in which concept manifolds untangle into linearly separable subspaces immediately before computation and compress afterward.
-
Conformal Thinking: Risk Control for Reasoning on a Compute Budget
Conformal risk control with upper and lower thresholds lets LLMs adaptively stop reasoning while guaranteeing a maximum error rate and minimizing token use.
-
Learning More from Less: Unlocking Internal Representations for Benchmark Compression
Using aligned hidden states from ten LLMs to select a small representative question set lets RepCore estimate full-benchmark accuracy and rankings more accurately than output-only methods.
-
Entropy After </Think> for reasoning model early exiting
Entropy After </Think> (EAT) enables early exiting in reasoning LLMs by tracking entropy stabilization after a </think> token, cutting token use 12-22% on MATH500 and AIME2025 with no accuracy loss.
-
Prefix-Safe Bayesian Belief Tracking for LLM Reasoning Reliability:Separating Calibration from Ranking
SBBT separates Brier-score calibration gains from AUROC ranking gains in prefix-conditioned success estimation for LLM math reasoning, with structure-aware signals yielding up to +0.110 AUROC over baselines.
-
When Self-Belief Misleads: Active Label Acquisition for Reinforcement Learning with Verifiable Rewards
RLAVR uses the Corrective Advantage Gap metric and CARE policy to actively acquire ground-truth labels for key samples, stabilizing RLVR training and boosting performance with limited annotation budgets.
-
TRACES: Tagging Reasoning Steps for Adaptive Cost-Efficient Early-Stopping
TRACES tags reasoning steps to enable adaptive early stopping, cutting token use by 20-50% on MATH500, GSM8K, AIME, MMLU and GPQA with comparable accuracy.
-
LLM Reasoning Is Latent, Not the Chain of Thought
LLM reasoning is primarily mediated by latent-state trajectories rather than by explicit surface chain-of-thought outputs.
-
How Long Reasoning Chains Influence LLMs' Judgment of Answer Factuality
Weak LLM judges accept wrong answers more often when shown fluent reasoning chains, while strong judges use them partially but still get misled by high-quality-looking but flawed reasoning.
-
Training-Free versus Training-Based Intent Classification in LLMs: Accuracy, Robustness, and Failure Modes
Statistical classifiers built on LLM activation norms and coordinates match or beat trained MLP heads on coarse intent routing and resist camouflage better, while MLPs win on fine-grained subfield distinctions.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.