PaperlayerView exact revision on arXiv

Source-linked AI summary

Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models

Jean de Dieu Nyandwi, Leena Mathur, Yonatan Bisk, Robert Hawkins, Graham Neubig

arXiv:2608.13760v1cs.CLcs.AIcs.CVcs.LG

TL;DR

The paper asks whether the reasoning behaviors amplified by thinking-oriented training are the behaviors most associated with correctness. Across models, modalities, and benchmarks, it finds an Amplification-Lift Gap: deliberative behaviors are amplified, while confidence calibration and other high-lift behaviors are not.

  • Problem

    Longer reasoning traces reveal which behaviors models use, but accuracy alone does not show which behaviors are associated with correctness or whether thinking training amplifies them.

  • Method

    The authors use Behavioral Lift and a cross-modal taxonomy to analyze nine reasoning behaviors across 15,282 traces from 15 models and six benchmarks.

  • Results

    The amplified behaviors are not the strongest correctness signals: confidence calibration has +72–80% lift, whereas uncertainty acknowledgment is strongly amplified but weakly or negatively associated with correctness.

  • Takeaways & Limitations

    Reasoning training and evaluation should reward calibrated, evidence-grounded behaviors rather than longer traces, backtracking, or uncertainty expressions alone.

  • Takeaways & Limitations

    Behavioral Lift is descriptive, so high-lift behaviors may correlate with correctness without causing better performance, and controlled training studies remain necessary.

Abstract

from arXiv · show

Which reasoning behaviors are associated with correct answers in reasoning models, and does reasoning-oriented training amplify those behaviors? This distinction is important because reasoning-oriented training can make traces look more deliberative without amplifying the behaviors most tied to model correctness. We quantify this mismatch with Behavioral Lift, a metric that measures how much correctness changes when a behavior is present versus absent in a model's reasoning trace. Across 15 models and 6 benchmarks spanning text-only and vision-language reasoning, we annotate 15,282 traces with a taxonomy whose core behaviors are defined for both LLM and VLM traces. We find evidence for an Amplification-Lift Gap, in which thinking models strongly amplify self-correction, hypothesis testing, and uncertainty acknowledgment, while the highest-lift behaviors are confidence calibration, knowledge alignment, and self-awareness. Confidence calibration is among the strongest positive signals of correctness in both modalities, yet is barely amplified; uncertainty acknowledgment is amplified by 3--7$\times$, yet is weakly or negatively associated with correctness. We find that reasoning-oriented training does not preferentially amplify the highest-Lift behaviors, motivating process-level objectives that reward calibrated and grounded reasoning rather than surface form alone.

1 Introduction

This paper asks whether thinking-oriented training amplifies reasoning behaviors associated with correctness, distinguishing behavioral prevalence from behavioral lift. Across language and vision-language models, it identifies an Amplification-Lift Gap: deliberative-looking behaviors are not necessarily the most predictive of success.

  • Main findings: Thinking models selectively amplify self-correction, hypothesis testing, and uncertainty acknowledgment while leaving confidence calibration and knowledge alignment largely unchanged.The paper reports that no behavior is both strongly amplified and strongly predictive of success.
  • Motivation and framing: Behavioral Lift separates how often a behavior appears from how strongly it is associated with reasoning correctness.The distinction addresses cases where behaviors become more frequent without becoming more diagnostic of successful reasoning.
  • Study design: 9 behaviors and 15,282 annotated traces from 15 LLMs and VLMs across 6 benchmarks support a cross-modal analysis.The taxonomy spans reasoning behaviors, failure modes, reasoning quality, reasoning types, summary metrics, and visual grounding for VLMs.
  • Main findings: Thinking-oriented models do not exhibit the highest-Lift behaviors, providing empirical evidence for an Amplification-Lift Gap.The gap reflects a disconnect between behaviors amplified by training and behaviors most associated with correctness.
  • Mechanism: Recovery helps explain thinking-model gains when tasks reward extended computation or recovery from failures.The paper identifies recovery as a mechanism behind gains from thinking-oriented models.

2 Taxonomy and Metrics

The paper introduces a cross-modal taxonomy of reasoning behaviors and failures, alongside metrics that separate behavioral presence from its association with correctness and quantify recovery despite failures. The taxonomy emphasizes process-level annotation, modality-neutral higher-order behaviors, and epistemic grounding.

  • Taxonomy: The taxonomy annotates reasoning processes alongside final answers, distinguishes valid logic from lucky guesses, and separates behavioral presence from Behavioral Lift.These principles are intended to distinguish reasoning quality from answer correctness alone and frequent behaviors from behaviors associated with success.
  • Taxonomy: The taxonomy contains six groups, with higher-order behaviors supporting cross-modal comparison and failure modes supporting Recovery Rate analysis.Other groups provide modality-specific grounding and descriptive context, including reasoning quality and reasoning types.
  • Taxonomy: Nine higher-order behaviors span both modalities and comprise control/regulation, monitoring/judgment, and epistemic grounding.They include planning, goal tracking, hypothesis testing, self-correction, uncertainty acknowledgment, confidence calibration, self-awareness, evidence citation, and knowledge alignment.
  • Metrics: Behavioral Lift measures how correctness differs when a behavior is present versus absent, with positive Lift indicating higher accuracy and negative Lift indicating lower accuracy.Lift is descriptive rather than causal and is computed on pooled samples across all models within each modality.
  • Metrics: Recovery Rate measures whether a model reaches correct answers despite detected reasoning failures, with a high rate indicating successful recovery despite failures.The relevant failure set includes all 7 failure modes for the modality.

3 Experimental Setup

The study evaluates matched thinking and instruction-tuned model variants across diverse reasoning and knowledge benchmarks, using standardized inference, trace annotation, and judge validation. It annotates 15,282 responses with behavior labels produced from full model outputs and ground-truth answers.

  • Models: 15 open-weight models spanning 3B–9B parameters include matched thinking and non-thinking variants across LLM and VLM families.The set contains 4 thinking and 3 non-thinking VLMs, plus 4 thinking and 4 non-thinking LLMs.
  • Benchmarks: Six benchmarks span visual, logical, mathematical, multimodal-knowledge, and knowledge-intensive tasks across three task types per modality.VLMs use VisualPuzzles, MathVista, and MMMU; LLMs use LogiQA2, MATH-500, and MMLU-Pro.
  • Data collection: 15,282 responses are annotated after targeting 350 responses per benchmark, except MathVista, which uses 300 testmini responses.Final counts fall slightly below targets because responses that fail output parsing are excluded.
  • Inference: All models are evaluated under standardized conditions with full reasoning traces retained for annotation.Additional inference details are provided in Appendix A.
  • Behavioral annotation: GPT-4o annotates each response using the question, ground-truth answer, and full model output to produce binary labels for relevant behaviors.For confidence calibration, annotations are true only when expressed certainty matches the observed strength of reasoning.

4 Results

Thinking training amplifies self-correction, hypothesis testing, and uncertainty acknowledgment, but not the behaviors most strongly associated with correctness. Confidence calibration, knowledge alignment, and self-awareness have the highest lifts, while recovery and task structure determine when thinking models help.

  • Behavior amplification: Self-correction, hypothesis testing, and uncertainty acknowledgment are amplified across all seven model families, both modalities, and all six benchmarks.Their prevalence is 21–55% versus 3–15%, 22–52% versus 4–18%, and 25–85% versus 4–28%, respectively, in thinking versus comparison models.
  • Behavioral Lift: Confidence calibration shows +72.2% Lift in VLMs and +79.6% in LLMs, yet thinking models are no more likely to exhibit it than instruct models.Accuracy is 98.8% versus 26.7% when present versus absent for VLMs, and 99.6% versus 20.0% for LLMs.
  • Behavioral Lift: Uncertainty acknowledgment has −16.1% Lift in VLMs and −13.9% in LLMs, while hypothesis testing has +1.0% Lift in both modalities.Self-correction has modest positive Lift: +20.1% in VLMs and +12.4% in LLMs.
  • Behavioral Lift: The Lift ranking is stable across modalities, benchmarks, controls, response-complexity strata, and per-model analyses, with confidence calibration, self-awareness, and knowledge alignment highest.Hypothesis testing and uncertainty acknowledgment rank lowest, and same-question and probing analyses suggest the ranking is not a question-difficulty or surface-labeling artifact.
  • Recovery and task dependence: Thinking models gain through fewer failures or better recovery, outperforming instruct models on computation-rewarding tasks but losing on LogiQA2, where shortcuts work.On LogiQA2, instruct models reach 58.4% versus 54.1% accuracy and recover at 24.5% versus 11.1%; self-correction after detected failures is associated with +39pp recovery.
  • Scaling: The amplification-Lift gap persists with scale: self-correction Lift falls from +30.0% at 2B to +5.9% at 32B, while confidence calibration Lift rises from +57.7% to +68.7%.At 32B, thinking models self-correct in 61.3% of responses versus 11.3% for instruct models, while confidence calibration reaches 78.0% for instruct versus 71.7% for thinking.

5 Discussion

The discussion argues that visible deliberation is not a reliable proxy for correctness: confidence calibration tracks reasoning strength, whereas uncertainty and other surface markers do not necessarily do so. It therefore motivates process objectives centered on grounded, calibrated reasoning while noting limitations of visible traces, automated labels, and descriptive Behavioral Lift.

  • Behavioral Lift distinguishes frequent deliberation from useful reasoning by comparing correctness when behaviors are present versus absent.Self-correction and uncertainty can signal either successful reasoning or earlier errors and confusion.
  • Confidence calibration tracks correctness by aligning expressed certainty with reasoning strength, unlike uncertainty acknowledgment, which merely expresses doubt or hesitation.Explicit uncertainty is informative only when it reflects the available evidence.
  • Visible deliberation is an unreliable proxy for reasoning quality because correctness-focused training can produce longer traces with more correction, search, and hesitation.Thinking helps most when tasks reward extended computation.
  • Process objectives should reward evidence-grounded claims, appropriate domain framing, recognition of underspecified information, and confidence calibrated to reasoning strength.Longer traces, backtracking, and uncertainty expressions are not objectives in themselves; Behavioral Lift can audit whether objectives reward success-associated behaviors.
  • The analysis is limited by potentially incomplete or unfaithful visible traces, possible automated-judge bias, and Behavioral Lift’s descriptive rather than causal interpretation.More controlled training studies are needed despite validation, robustness checks, same-question controls, probing analyses, and prompting results consistent with the ranking.

6 Related Work

Related work spans reasoning-oriented training that elicits long chains of thought, behavioral analyses of reasoning traces, and process-level supervision and evaluation beyond final-answer accuracy.

  • Reasoning-oriented training and long chain-of-thought: Reasoning-oriented training uses RL or hybrid post-training to elicit long traces featuring behaviors such as backtracking and self-correction.Studies also examine how long chain-of-thought reasoning emerges during training and report that some core abilities are already present in base models.
  • Behavioral analysis of reasoning traces: Behavioral-trace research applies cognitive taxonomies, strategy discovery, clustering, steering, semantic-flow analysis, and latent-computation perspectives.These studies investigate which cognitive behaviors characterize traces and how reasoning strategies can be identified or influenced.
  • Process supervision and evaluation beyond accuracy: Process supervision and evaluation assess reasoning at the step level and test whether models detect errors or avoid shortcuts beyond producing correct final answers.Chain-of-thought prompting established that explicit reasoning can improve final-answer accuracy, while process reward models and verifier-style benchmarks evaluate intermediate reasoning.

7 Conclusion

Across 15,282 traces from 15 models and 6 benchmarks, thinking training consistently amplifies self-correction, hypothesis testing, and uncertainty acknowledgment. However, these amplified behaviors are not the ones most associated with correctness: confidence calibration is a stronger positive signal, while several higher-lift behaviors remain largely unchanged.

  • Conclusion: 15,282 traces from 15 models and 6 benchmarks quantify the behavioral effects of thinking training.The analysis spans the paper’s full set of annotated traces, models, and benchmarks.
  • Conclusion: Thinking training consistently amplifies self-correction, hypothesis testing, and uncertainty acknowledgment.These are the behaviors whose presence increases most consistently after thinking-oriented training.
  • Conclusion: Confidence calibration is one of the strongest positive signals of correctness, whereas amplified behaviors are not the ones most associated with correctness.The conclusion identifies a mismatch between behavioral amplification and association with correct answers.
  • Conclusion: Several higher-lift behaviors remain largely unchanged under thinking training.Training does not consistently increase the behaviors most associated with correctness.

Ethics Statement · A Inference Details · B Qualitative samples and surface markers of annotated behaviors

The paper studies reasoning behavior observationally through automated annotation of benchmark outputs, while acknowledging that visible traces may not fully reflect internal reasoning. Evaluations use benchmark-native, mostly zero-shot prompting and model-family decoding defaults, with qualitative examples illustrating how annotated behaviors appear in model outputs.

  • Ethics Statement: The analysis is observational, focuses on benchmark responses rather than real-world decisions, and does not claim that visible traces fully reflect internal reasoning.Prompts, metrics code, and annotations will be released to support transparency and reproducibility.
  • Ethics Statement: LLMs assisted with grammar, editing, implementation, and verification of selected analyses, while LLM-based annotation is documented in the methodology.
  • A Inference Details: LLMs are evaluated with lm-evaluation-harness and VLMs with lmms-eval, using benchmark-native wrappers and zero-shot prompting unless otherwise noted.CoT zero-shot variants are used when available, and instruct-model prompts follow benchmark or evaluation-suite templates.
  • A Inference Details: Reasoning-enabled LLMs use temperature 0.6, top-p 0.95, top-k 20, and min-p 0, whereas Qwen2.5 instruct models use temperature 0.7 and top-p 0.8.Nemotron-Base is evaluated with greedy decoding at temperature 0.
  • A Inference Details: Qwen3-VL-8B and GLM-4.1V-9B thinking variants use temperature 1.0, top-p 0.95, top-k 20, and a maximum generation length of 40,960 tokens.Their instruct variants use temperature 0.7, top-p 0.8, top-k 20, presence penalty 1.5, and a maximum generation length of 16,384 tokens.
  • A Inference Details: The evaluations retain model-specific defaults and benchmark wrappers and add no extra prompting intended to elicit particular behaviors.Table 3 lists the decoding settings used for each model family.
  • B Qualitative samples and surface markers of annotated behaviors: Behavior-positive traces make annotation labels auditable by showing distinct surface forms, including overlap at label boundaries.
  • B Qualitative samples and surface markers of annotated behaviors: Uncertainty acknowledgment uses hedging, self-awareness uses information audits, self-correction uses explicit reasoning breaks, and hypothesis testing uses branching language.Examples include “I’m not sure,” “I’m confused,” “wait,” “actually,” “let’s start over,” “alternatively,” and “suppose.”

C Robustness to annotation noise and question difficulty

Robustness analyses show that Behavioral Lift rankings persist after controlling within questions and perturbing annotations, while calibration tracks sound reasoning beyond answer correctness alone. Behaviors generally appear before answer commitment, but within-question comparisons do not establish causal direction.

  • Within-question analysis: Within-question comparisons reproduce the broad behavior ranking, constraining question-difficulty explanations without resolving causal direction.Traces are compared within the same model and MATH-500 question, but latent reasoning-path quality remains uncontrolled.
  • Sensitivity to random annotation noise: 5%, 10%, 15%, and 20% random label flips leave the main Behavioral Lift ranking stable across 1000 trials.For LLMs, confidence calibration and knowledge alignment remain top-ranked, while uncertainty acknowledgment remains lowest-ranked in every trial.
  • Lucky-guess analysis: Calibration is far less common on lucky guesses than non-lucky responses across LLMs, VLMs, thinking models, and instruct models.This supports calibration as a marker of sound reasoning rather than final-answer correctness alone.
  • Temporal position of behaviors: 91–99% of traces show all four behaviors before the answer, with mean behavior positions at 27–35% versus 79–94% for answer commitment.The shared timing pattern reflects the general structure of reasoning traces rather than calibration specifically.

D Length-Controlled Prevalence Analysis

The analysis tests whether longer thinking-oriented traces explain their higher prevalence of self-correction, hypothesis testing, and uncertainty acknowledgment. Length explains part of the pattern, particularly in the shortest traces, but does not eliminate it.

  • Length-Controlled Prevalence Analysis: Responses are grouped into shared word-count quintiles within each modality to recompute thinking-minus-comparison prevalence gaps.This controls for trace length when evaluating amplification patterns.
  • Length-Controlled Prevalence Analysis: Length accounts for part of the amplification pattern, especially among the shortest traces, but does not eliminate it.The residual pattern remains after comparing traces within matched length bins.

E Linear Probing of Behavioral Representations · F Prompting Behavioral Lift at Inference Time

Linear probes show that several annotated behaviors are internally decodable, with thinking-model decodability aligned with Behavioral Lift especially on incorrect traces. At inference time, prompting high-Lift behaviors improves accuracy relative to baseline, whereas pervasive uncertainty prompting consistently harms it, though the preliminary design does not isolate individual behaviors.

  • E Linear Probing of Behavioral Representations: Teacher-forced layer-35 hidden states support logistic-regression probes with question-grouped cross-validation, testing whether annotations reflect internal structure rather than surface labels.The probes use generated portions of saved traces and prevent question-level leakage.
  • E Linear Probing of Behavioral Representations: ρ = +0.86 (p = 0.014) is the thinking model’s incorrect-trace rank correlation between probe decodability and Behavioral Lift.Knowledge alignment and goal tracking are among the most separable behaviors in incorrect thinking traces.
  • E Linear Probing of Behavioral Representations: ρ = −0.71 (p = 0.071) is the instruct model’s incorrect-trace correlation, where hypothesis testing and uncertainty acknowledgment are among the most decodable behaviors.Confidence calibration is nearly absent from incorrect traces in both models: 1.7% thinking and 0.7% instruct.
  • E Linear Probing of Behavioral Representations: Several behaviors, including confidence calibration and knowledge alignment, are decodable from hidden states, supporting the annotation labels and their behavioral interpretation.In the thinking model, decodability is directionally aligned with Behavioral Lift, especially on incorrect traces.
  • F Prompting Behavioral Lift at Inference Time: High-Lift prompting encourages confidence calibration, knowledge alignment, and self-awareness, while Low-Lift prompting encourages pervasive uncertainty acknowledgment across MATH-500, MMLU-Pro Math, and GPQA.The conditions use identical decoding settings on 500, 550, and 448 questions, respectively.
  • F Prompting Behavioral Lift at Inference Time: +0.4 to +4.4pp confidence calibration and +3.8 to +8.0pp knowledge alignment accompany High-Lift prompting, while Low-Lift uncertainty acknowledgment reaches 94.4–99.3%.High-Lift prompting reduces uncertainty acknowledgment by −18.1 to −21.2pp; Low-Lift reduces confidence calibration by up to 24.3pp.
  • F Prompting Behavioral Lift at Inference Time: Within response-complexity bins C2–C4, confidence calibration remains strongly positive and uncertainty acknowledgment remains negative in both modalities.This supports the claim that the amplification–Lift disconnect is not explained by response length or problem difficulty.

G Visual Claim Accuracy Matters More Than Image References

For VLM traces, accurate visual claims are strongly associated with correctness, whereas merely mentioning image content is not. Thinking models reduce visual neglect but increase visual hallucination, while confidence calibration remains informative across grounding quality.

  • Visual Claim Accuracy: Visual references and visual claim accuracy separate image mention from correct image use in VLM traces.The annotations record whether traces mention concrete image content and whether those statements are factually correct.
  • Visual Claim Accuracy: Visual claim accuracy has +60.9 Behavioral Lift, with correct responses occurring 87.8% of the time versus 26.9% when claims are inaccurate.Simply mentioning visual content has -3.3 Lift; accuracy remains positive on VisualPuzzles (+71.7), MathVista (+54.3), and MMMU (+45.1).
  • Visual Failures: Thinking models show less visual neglect than instruct models overall, 36.0% versus 47.4%, but more visual hallucination, 24.7% versus 18.0%.Longer traces appear to attend to images more actively while creating more opportunities for incorrect visual claims.
  • Visual Claim Accuracy: Confidence calibration has +43.9 Lift among traces with accurate visual claims and +73.5 Lift among traces with inaccurate claims.Calibration therefore remains strongly associated with correctness after conditioning on visual grounding quality.
Loading 2608.13760v1…