- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes LoCo, a reasoning compression method that constructs preference pairs at selected intermediate reasoning states. It first detects candidate overthinking points using model confidence and confidence variation, verifies them with a stronger judge, and then regenerates shorter correct continuations from the same frozen prefixes. These local preference pairs are used for DPO training.
LoCo is trained on mathematical traces from DeepSeek-R1-Distill-Qwen-7B and evaluated on five math and three code benchmarks. It substantially reduces completion length while maintaining or improving accuracy. The paper also studies global versus local preference construction, external versus self-generated continuations, and different cutpoint strategies.
Critical strengths
- Strong compression results. LoCo reduces completion length by about 42%–45% on the math benchmarks while maintaining or improving average accuracy. The gains are also consistent in AUC, giving a strong accuracy–efficiency trade-off.
- Cross-domain transfer. Although LoCo is trained only on mathematical traces, it also reduces completion length on all three code benchmarks without hurting average accuracy relative to the base model. This suggests that the learned compression behavior is not limited to the training domain.
Critical weaknesses
- Missing comparison with the closest prior methods. The paper discusses SSPO and ReCUT as closely related step-level compression methods, but does not compare against them experimentally. The current comparisons with DAST and TokenSqueeze do not clearly isolate the benefit of LoCo’s specific local preference construction.
- The cutpoint ablation does not fully validate the overthinking detector. LoCo outperforms RandCut and UniCut, but its selected cutpoints are also later in the trajectory on average. This leaves open whether the gain comes from detecting overthinking states or simply from compressing at later, safer positions. A position-matched random baseline would be important.
- Evaluation is limited to a single reasoning backbone. All LoCo experiments are conducted on DeepSeek-R1-Distill-Qwen-7B. Since the detector and its thresholds are calibrated from the model’s own confidence statistics, it is unclear whether the method generalizes across model scales or to stronger and more recent reasoning models.
Questions for authors
-
How does LoCo compare with the closest step-level compression methods, especially SSPO and ReCUT, under a comparable backbone and training setup?
-
Does LoCo still outperform a random-cut baseline whose cutpoint-position distribution is matched to LoCo’s detected states?
-
Could the authors report basic evaluation of the overthinking detector itself, such as the fraction of screened states retained by the judge and the accuracy of selected states on a manually checked subset?
-
Have the authors tested LoCo on other reasoning backbones, especially different model scales or stronger, more recent models?
3 comments
Author Response to Reviewer bXD1 (Part II)
4. Generalization to Qwen3-8B (Q4)
We further reproduce the full LoCo pipeline on Qwen3-8B (thinking mode), a stronger and more recent reasoning model. We recalibrate the confidence and volatility thresholds using Qwen3-8B’s own traces and leave the remaining pipeline unchanged.
| Model | MATH500 Acc | MATH500 Token | MATH500 AUC | AMC23 Acc | AMC23 Token | AMC23 AUC | Olympiad Acc | Olympiad Token | Olympiad AUC |
|---|---|---|---|---|---|---|---|---|---|
| Qwen3-8B | 95.22±0.35 | 5501±94 | 80.38±0.26 | 94.06±3.76 | 9132±573 | 71.09±2.44 | 69.56±0.66 | 11333±117 | 50.81±0.40 |
| LoCo-Qwen3-8B | 95.25±0.37 (+0.03) | 3433±60 (↓37.6%) | 86.29±0.38 (+5.91) | 95.94±1.86 (+1.88) | 5272±326 (↓42.3%) | 82.10±1.67 (+11.01) | 68.57±0.70 (−0.99) | 8332±127 (↓26.5%) | 56.65±0.45 (+5.84) |
| Model | AIME24 Acc | AIME24 Token | AIME24 AUC | AIME25 Acc | AIME25 Token | AIME25 AUC |
|---|---|---|---|---|---|---|
| Qwen3-8B | 72.50±4.96 | 15082±940 | 46.32±2.04 | 63.75±5.18 | 18221±1011 | 37.79±3.16 |
| LoCo-Qwen3-8B | 73.33±2.52 (+0.83) | 10735±469 (↓28.8%) | 54.31±1.87 (+7.99) | 66.67±2.52 (+2.92) | 13629±369 (↓25.2%) | 47.05±1.53 (+9.26) |
The results demonstrate that LoCo transfers beyond the original R1-Distill-Qwen-7B backbone.
Gentle Reminder
Dear Reviewer,
We hope this message finds you well. We have tried our best to address your concerns. Should there be any concerns or points you wish us to clarify further? Please do not hesitate to reach out. We are more than willing to extend our conversation and eagerly anticipate any further discussions that may arise.
Best regards,
All Authors
Author Response to Reviewer bXD1 (Part I)
We thank the reviewer for the careful and constructive comments. Below, we clarify the comparison with prior methods and provide additional ablation and generalization evidence for the detector and across model families. We have revised the paper in response to the comments. For the reviewers' convenience, newly added or revised content is shaded in light blue.
1. Comparison with ReCUT and SSPO (W1, Q1)
Beyond the detector, LoCo differs from ReCUT and SSPO in two key aspects.
a. Scalability to long-reasoning models. ReCUT and SSPO are developed and evaluated on instruction-tuned models (e.g., Qwen2.5-7B-Instruct), where CoT is elicited through prompting, rather than on dedicated long-reasoning models such as DeepSeek-R1-Distill-Qwen-7B. ReCUT repeatedly expands long and short continuations at each reasoning step, while SSPO requires dense step-wise value estimation and online RL. Both introduce substantial overhead when applied to long reasoning traces. In contrast, LoCo identifies candidate states with a single forward pass over existing traces, regenerates continuations only for selected states, and performs offline DPO, resulting in a more efficient and scalable training pipeline.
b. State-conditioned stopping supervision. ReCUT uses step-wise exploration during data construction, but its final DPO objective still compares complete trajectories. SSPO leverages step-wise values to assess whether each reasoning step progresses toward the correct answer. However, answer probability may temporarily decrease during necessary exploration, making a value drop an unreliable signal for stopping. In contrast, LoCo compares two correct continuations from the same frozen reasoning state, directly isolating the preference between continuing and concluding.
Neither ReCUT nor SSPO reports results on R1-Distill backbones, and no public implementation is available for this setting. We therefore compare against DAST and TokenSqueeze under the same backbone and evaluation protocol. To isolate LoCo’s local preference construction specifically, the position-matched ablation below provide more direct controlled comparisons.
2. Position-Matched Control (W2, Q2)
To evaluate cut-point selection, we follow the reviewer's suggestion and randomly sample cut points from the same relative-position distribution as our detector, isolating the effect of detector selection from cut-point location.
With cut-point locations held constant, replacing LoCo's detector with random cuts reduces accuracy by 1.87 / 4.17 / 4.59 points and AUC by 1.35 / 3.16 / 3.05 points on AMC23 / AIME24 / AIME25, respectively. LoCo also achieves fewer tokens, demonstrating that its gains come from more effective overthinking detection.
3. Evaluation of the Detector (Q3)
We thank the reviewer for this suggestion. We further compare our detector with a high-confidence-only baseline using the same judge model. The detection accuracy results are reported below:
The proposed detector produces substantially fewer but more precise candidates than a high-confidence-only rule. We additionally provide visualizations of representative reasoning traces, with the corresponding step-level confidence mean and variance curves, as well as the position where the correct answer first appears (Appendix B.1). The visualizations show that our detector effectively identifies overthinking states after the correct answer emerges. In contrast, high confidence alone leads to more false positives, as confidence does not necessarily indicate that reasoning is complete.