Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models
Jean de Dieu Nyandwi, Leena Mathur, Yonatan Bisk, Robert Hawkins, Graham Neubig
COLM 2026
cs.CL, cs.AI, cs.CV, cs.LG
2026-08-14
Behavioral analysis of 15,282 traces across 15 models shows thinking training amplifies self-correction, hypothesis testing, and uncertainty acknowledgment, while the behaviors most associated with correctness are confidence calibration, knowledge alignment, and self-awareness; thinking models win through 2-3x higher recovery from mid-stream failures instead.
Thinking models produce chains of thought that look deliberative: they self-correct, test hypotheses, and acknowledge uncertainty. But "looks deliberative" and "the way it thinks is tied to being right" are different claims. Reasoning-oriented training may amplify surface behaviors, lengthening and solemnifying traces, without amplifying the behaviors actually associated with correctness. This paper separates "amplified" from "predictive", measures both across a large model and benchmark grid, and finds a real gap.
Two core quantities. Prevalence: how often a behavior appears in a trace. Behavioral Lift: the probability of a correct answer when the behavior is present minus the probability when it is absent, Lift(b) = P(correct|b) minus P(correct|not b). Positive lift means the behavior tracks higher accuracy; negative lift means the opposite.
The sample: 15 open-weight models (3B-9B; 7 VLMs and 8 LLMs, with same-family thinking/instruct pairs wherever available), 6 benchmarks (VisualPuzzles, MathVista, and MMMU for VLMs; LogiQA2, MATH-500, and MMLU-Pro for LLMs; roughly 350 samples each), and 15,282 annotated traces. Annotation uses GPT-4o as judge over nine cross-modal higher-order behaviors: planning, goal tracking, hypothesis testing, self-correction, uncertainty acknowledgment, confidence calibration, self-awareness, evidence citation, and knowledge alignment, plus seven failure modes per modality (logical failure, post-hoc rationalization, shortcut, lucky guess, and modality-specific others). Judge reliability was checked against three independent judges (DeepSeek-V3, Gemini-2.5-Flash, Gemini-3-Flash) on 600 stratified samples, with kappa 0.51-0.82 on focal behaviors; a manual audit of 120 traces and 720 binary decisions agreed with GPT-4o 95.1% of the time (kappa 0.902).
Three behaviors are strongly amplified by thinking training: self-correction appears in 21-55% of thinking-model responses versus 3-15% for instruct models; hypothesis testing 22-52% versus 4-18%; uncertainty acknowledgment 25-85% versus 4-28%, a 3-7x multiplier.
The Lift ranking runs the other way:
| Behavior | Lift (VLM) | Lift (LLM) | Prevalence change |
| Confidence calibration | +72.2% | +79.6% | roughly unchanged, instruct slightly higher on several benchmarks |
| Knowledge alignment | +53.7% | +80.3% | not amplified |
| Self-awareness | +62.0% | +52.7% | not amplified |
| Self-correction | +20.1% | +12.4% | amplified 4-17x |
| Hypothesis testing | +1.0% | +1.0% | amplified 2-13x |
| Uncertainty acknowledgment | -16.1% | -13.9% | amplified 3-7x |
Confidence calibration is the single strongest correctness signal: when present, VLMs answer 98.8% correctly versus 26.7% when absent; LLMs 99.6% versus 20.0%. It appears in 94.7% (VLM) and 95.9% (LLM) of sound-reasoning traces but only 7.6% and 3.4% of lucky guesses, so it tracks reasoning quality rather than final-answer luck. Uncertainty acknowledgment, by contrast, is more common in incorrect traces (30.8% versus 20.1%) and shows up in 56.7% of post-hoc rationalization traces.
So why do thinking models score higher? Recovery. The rate of reaching a correct answer after at least one detected failure is 2-3x higher for thinking models on computation-heavy benchmarks: VisualPuzzles 23.0% versus 8.4% (2.7x), MATH-500 40.8% versus 17.8% (2.3x), MMLU-Pro 16.8% versus 6.3% (2.7x). On pattern-matching tasks the pattern flips: on LogiQA2, instruct models recover better (24.5% versus 11.1%).
The gap does not close with scale. At 32B, thinking models self-correct in 61.3% of responses versus 11.3% for instruct, but 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%. Larger thinking models acknowledge uncertainty less (66.7% down to 43.7%). Seven frontier closed models across five providers show the same ranking on GPQA-Diamond: knowledge alignment +84.2, confidence calibration +77.9, uncertainty acknowledgment -23.2.
A preliminary prompting experiment points the same way. Injecting high-Lift behavior prompts (calibration plus grounding) into OLMo-3-7B-Think improved MATH-500 accuracy by 5.8pp and recovery from 57.1% to 68.9%; on MMLU-Pro Math the high-versus-low prompt gap reached 15.7pp. Injecting a low-Lift prompt that saturates uncertainty acknowledgment degraded accuracy on all three benchmarks. The authors flag this as preliminary: the prompts change several properties at once and the GPQA differences are not statistically significant.
For training: if the goal is reasoning quality, process-level objectives should target calibration, grounding, and self-awareness, the high-Lift behaviors, rather than rewarding traces for containing self-corrections or hypothesis tests. Surface form is exactly what optimizers exploit; rewarding it rewards performance.
For evaluation: a long chain of thought, or a chain of thought containing self-corrections, is not evidence of reasoning quality. Verbal uncertainty is not calibration either; it tracks task difficulty.
For usage: on pattern-matching tasks thinking models can lose ground (the LogiQA2 recovery reversal), while their biggest edge is on tasks where mid-stream errors need to be caught and repaired.
The authors list four: the analysis covers visible traces only, and written reasoning may be incomplete, post hoc, or unfaithful to internal computation; labels come from automated judges, so systematic bias remains possible despite multi-judge validation and manual audits; Lift is descriptive, and high-lift behaviors may be consequences rather than causes, or co-occur with a third useful property; the prompting experiment is preliminary and does not isolate single behaviors. This reader adds one more: Lift magnitudes are naturally coupled to prevalence. A behavior present in 95% of sound traces compares against a tiny complement, so numbers like 99.6% versus 20.0% should be read as conditional-probability gaps under extreme imbalance, not as linear causal effects. The model grid is 3B-9B plus one frontier validation; the middle of the scale is empty.