Conserved Verdicts, Drifting Commitments: Identifying Harmful Overthinking
Abstract
A reasoning model that has reached a correct answer can lose it by continuing to think. Deciding whether to stop means comparing two probabilities at the same point of a chain: that the answer the model would give if stopped now is correct (its commitment), and that the answer it will finish with is correct (its verdict). We show that the verdict alone, which is what final-correctness probes estimate, cannot make this comparison. As a forecast of a fixed outcome it has no predictable drift, whether continuing helps or hurts. The commitment-verdict gap equals the expected remaining change in answer quality and the accuracy gained by stopping. A second identity splits any accuracy-versus-budget curve into answer change within chains and change in which chains are still running. We measure both on MATH-500. At all 379 points where DeepSeek-R1-Distill-Qwen-1.5B holds a correct answer, selected without regard to how its chain ends, continuing loses 3.6 points on average (95% CI [2.2, 5.2]), and the loss concentrates in a few problems. In six reasoning models from three families, on MATH-500 and OlympiadBench and with budgets up to 16,384 tokens, declining accuracy curves come from which chains remain, while the chains still reasoning keep or improve their answers. An oracle stopping rule gains at most +1.8 and +2.5 points at 1.5B and 7B. The framework says what to measure before building a stopping rule.
Figures & tables
| Model | Setting | Survivor change | Within-cohort | Composition | |
|---|---|---|---|---|---|
| R1-Qwen 1.5B | FP16, levels 3–5 | 367 | |||
| R1-Qwen 1.5B | FP16, levels 4–5 | ||||
| R1-Qwen 1.5B | NF4, levels 4–5 | ||||
| R1-Qwen 7B | NF4, levels 4–5 | 262 | |||
| R1-Qwen 7B | 8-bit, levels 4–5 | ||||
| Qwen3 1.7B | FP16, levels 3–5 | 367 |
| R1-Distill-Qwen 1.5B | R1-Distill-Qwen 7B | |||
|---|---|---|---|---|
| Policy | Accuracy | Tokens | Accuracy | Tokens |
| Full budget | ||||
| Oracle stopping (gold labels) | ||||
| Drift alarm (ours) | ||||
| Checkpoint vote | ||||
| Answer agreement | ||||
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Pre-registered | outcome |
|---|---|
| Gate 0: apparatus calibration | passed (slope , ; matched floor ) |
| Gate 1(a): overthink-wrong chains | failed ( ) committed 7B escalation: |
| Gate 2: recovery at some | not met: no tested alarm recovers accuracy; this does not exclude local adverse drift |
| Verdict rule (viii) | not triggered: verdict AUC / does not clear (App. H ) |
| Robustness | outcome |
| Forcing operator | of truncations, yield no natural answer; on the rest, agreement ; greedy vs. sampled pts |
| Case/control | Level | Checkpoint | Case correct | Interval | Control correct | Interval |
|---|---|---|---|---|---|---|
| 9/109 | 5 | 2048 | 1/8 | 8/8 | ||
| 101/63 | 5 | 1536 | 6/8 | 8/8 | ||
| 120/57 | 2 | 1536 | 7/8 | 8/8 | ||
| 126/461 | 4 | 1024 | 2/8 | 6/8 | ||
| 128/471 | 3 | 512 | 0/8 | 8/8 | ||
| 238/459 | 5 | 1536 | 7/8 | 8/8 |
| 1.5B | 7B | |||||
|---|---|---|---|---|---|---|
| Refit | C | V | Gap | C | V | Gap |
| 1 | 0.2201 | 0.2332 | -0.0130 | 0.2254 | 0.2342 | -0.0088 |
| 2 | -0.6384 | 0.2512 | -0.8896 | 0.1837 | 0.2427 | -0.0590 |
| 3 | 0.2372 | 0.2553 | -0.0180 | 0.2403 | 0.2161 | 0.0242 |
| 4 | 0.3045 | 0.2790 | 0.0255 | 0.2193 | 0.2546 | -0.0353 |
| 5 | 0.2449 | 0.2539 | -0.0090 | 0.2161 | 0.2437 | -0.0276 |
| R1-Qwen 1.5B | R1-Qwen 7B | |
|---|---|---|
| Commitment score | ||
| Verdict score | ||
| Commitment minus verdict |
| R1-Distill-Qwen 1.5B | R1-Distill-Qwen 7B | |||
|---|---|---|---|---|
| Policy | Accuracy | Tokens | Accuracy | Tokens |
| Full budget | ||||
| Drift alarm | ||||
| Checkpoint vote | ||||
| Answer agreement | ||||
| Analysis | Specified | Reported in |
|---|---|---|
| Early-correct/final-wrong prevalence (Gate 1a), 7B escalation, predictable commitment drift, stopping-rule replays | before results | Secs. 4.2 , 4.4 ; App. H |
| Eight-sample resampling, temperature 0.9, 256-token grid, greedy vs. sampled forcing, Qwen3-1.7B, self-consistency | follow-up, before its run | Sec. 4.3 ; App. H.1 |
| R1-Distill-Llama-8B and Phi-4-mini-reasoning audits | follow-up, before its run | Table 1 |
| Survivorship decomposition; matched-prefix continuations; probe refits; cumulative-history ablation | follow-up after primary results; continuation and refit designs fixed before their runs | Secs. 4.2 – 4.3 ; Apps. J , K |
| Outcome-blind continuations, generic doubt, misleading hint, 16k budget, 1.5B in NF4, 7B in 8-bit, OlympiadBench, Qwen3.5-2B, problem concentration, difficulty match, oracle headroom | during review, before its run | Secs. 4.2 – 4.4 ; this appendix |
Overthinking in Language Models
Overthinking in language models studies when extra reasoning becomes redundant or harmful, and how to control it without losing accuracy.
See all 74 papers