acceptodds
Under review as a conference paper at ICLR 2027

Don't Waste Failures: Leveraging Failure Length Signals for Discriminative Thresholds in Efficient Reasoning

Abstract

Large reasoning models (LRMs) improve problem-solving accuracy through extended chains of thought, but can incur substantial cost from overly long reasoning. Existing reward-based approaches to reasoning efficiency often set length penalties using overall or success-only length statistics, failing to carefully differentiate how reasoning lengths differ between correct and incorrect rollouts. We show that leveraging this difference informs the design of length control: When failures are shorter than correct rollouts, we retain a higher penalty threshold, whereas longer failures motivate an earlier penalty. Based on this insight, we introduce OASIS (Outcome-Aware threshold Selection with Integrated Signals). OASIS estimates the reasoning-length threshold from the length distributions of correct and incorrect rollouts, determining where the length penalty begins. We further introduce a soft directional estimator to capture short and long failures separately, increasing the threshold for the former and decreasing it for the latter. Because a higher threshold leaves more reasoning unpenalized and weakens the effective length penalty, OASIS rescales the penalty gain to compensate for this weakening of the compression pressure while leaving the selected penalty-free region unchanged. Integrated into a length-aware GRPO reward, OASIS effectively reduces reasoning length relative to the base model across four base models evaluated on MATH500, AIME 24–26, and AMC23, while largely maintaining accuracy and yielding an average improvement of 1.7% points. Notably, on DeepSeek-R1-Distill-Qwen-1.5B, OASIS improves average accuracy by 4.5% points while reducing average reasoning length by 78.2%.

Primary area:
reinforcement learning
Keywords:
Efficient Reasoning, Reinforcement Learning, Length-aware reward shaping, Group Relative Policy Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7J81OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OASIS, which uses the length statistics of correct and incorrect rollouts for the same prompt to adjust the penalty threshold for correct responses during length-aware GRPO training. The authors construct candidates from the overall incorrect distribution and soft short/long failure components, select a threshold through scoring and curvature checks, and apply a batch-level iso-pressure gain to compensate for the penalty reduction caused by upward threshold shifts.

The authors evaluate OASIS on four reasoning models and five mathematical benchmarks. Relative to the base models, average generation length decreases by 32.7% and average accuracy increases by approximately 1.7 percentage points, with gains varying across models. The paper also includes SmartThinker retrained under matched settings and component ablations.

Critical strengths
  1. The paper provides a concrete analysis of existing length estimation. Under single-Gaussian moment matching and the stated variance conditions, the authors show how marginal pooling attenuates the threshold correction derived from failures. This offers a useful diagnosis of existing probabilistic length estimation; the directional candidates further seek to preserve distinct short- and long-failure structures.

  2. The matched comparison with the closest method supports some incremental value. Table 2 retrains SmartThinker on all four models with matched rollout settings and training steps. The final OASIS variant improves average accuracy and AE over this baseline on each model. Although it uses more tokens on two models, these results support an improved accuracy–length trade-off.

Critical weaknesses
  1. The method definition and ablation attribution are unclear. Section 4.1, Eqs. (9)–(12), uses a common mixture objective, whereas Algorithm 1, lines 43–46, uses different references by candidate type. The regular branch also leaves the subsequently used overall incorrect reference undefined. Table 3 does not specify whether adding directional candidates also changes the scoring density, making their independent contribution difficult to verify. These issues hinder reproduction and ablation attribution, but do not establish that the implementation is incorrect.

  2. The relative benefits of the added mechanisms remain insufficiently established. SmartThinker already provides Gaussian target lengths, penalties for overlong correct rollouts, and dynamic scaling. This paper adds incorrect-conditioned and directional candidates, along with batch-level restoration; the introduction's success-only characterization of SmartThinker is inaccurate. The matched comparison is valuable, but comparisons with cited representative methods such as ARLCP and DRPO are missing; ARLCP also shares the backbones and training data. Although iso has an on/off ablation, its comparator fixes g=1, leaving stronger penalties as an alternative explanation for its gains. The specific value of the restoration rule therefore needs a control comparison.

Questions for authors
  1. Does Table 3 keep the scoring density unchanged when adding directional candidates? Please reconcile the common mixture score with Algorithm 1's branch-specific scores and specify the implementation used.

  2. Could you compare iso against several moderately higher fixed gains under the same threshold-selection rule, to distinguish restoration from general penalty strengthening, and report variability across training seeds?

  3. Could you add comparable evaluations of representative methods such as ARLCP or DRPO and SmartThinker accuracy–length curves across penalty strengths? Please use a consistent evaluation protocol and clarify training-budget and optimizer differences.

I would reconsider my current score if the implementation rule is clarified and the added mechanisms demonstrate consistent gains at comparable cost.

4 comments

We thank for your careful review and constructive feedback. We address each concern below.


Weakness1: Clarification on Scoring Density and Score Reconciliation (Algorithm 1)

We appreciate the reviewer for catching this apparent discrepancy. While transcribing the method into pseudocode, the distinction between the derivation nulls (for all candidates) was mistakenly carried over into the J evaluation. The actual implementation, as the reviewer notes, and as stated in the body, scores all candidates against one common reference: the two-component incorrect mixture when the soft split is defined, a single Gaussian otherwise. We have corrected Algorithm 1 accordingly.

Consequently, the concern that adding directional candidates might alter the scoring density is resolved, as the scoring density remains strictly unchanged across all candidates. The gains reported in Table 3 therefore reflect the isolated effect of candidate expansion, confirming its clear independent contribution to identifying an elaborate threshold position.

Weakness2 : Misleading explanation about other methods and insufficient comparison.

We appreciate the reviewer's observation that characterizing SmartThinker simply as "success-only" in the Introduction was somewhat coarse. SmartThinker derives its target length from both correct rollouts and the marginal (all-rollout) distribution , which implicitly incorporates incorrect rollouts. Our intended distinction, which is formally analyzed in Section 4.1 (Eq. 4), is not between success-only vs. failure-aware, but between marginal pooling (SmartThinker) and explicit failure conditioning (OASIS). As derived in Eq. (4), fitting a single Gaussian to the pooled marginal distribution p(L ∣ x) introduces the attenuation factor κ = R/(R + π_C D²), which structurally absorbs the class separation between correct and incorrect rollouts into inflated variance.

We explicitly conditions on the incorrect distribution and models directional failure modes (Eqs. 5–9), eliminating this structural attenuation.

While this distinction is already rigorous in Section 4.1, we could slightly polish the introductory phrasing to prevent any potential ambiguity.

The comparison and iso pressure rule ablation will be further answered below.

Question1: Clarification on Scoring Density (Table 3) and Score Reconciliation (Algorithm 1)

Regarding the pseudocode in algorithm1, we appreciate the reviewer for catching this discrepancy. While transcribing the method into pseudocode, the distinction between the derivation nulls (for all candidates) was mistakenly carried over into the JJ evaluation. The actual implementation, as the reviewer notes, and as stated in the body, scores all candidates against one common reference: the two-component incorrect mixture when the soft split is defined, a single Gaussian otherwise. We have corrected Algorithm 1 accordingly.

Consequently, the concern that adding directional candidates might alter the scoring density is resolved, as the scoring density remains strictly unchanged across all candidates. The gains reported in Table 3 therefore reflect the isolated effect of candidate expansion, confirming its clear independent contribution to identifying an elaborate threshold position.

Question2: Iso-Pressure Restoration vs. Fixed Gain (Ablation Control)

Table: Accuracy and reasoning efficiency across iso-pressure settings. All experiments use DeepSeek-R1-Distill-Qwen-1.5B as the base model. We report accuracy (Acc.) and accuracy-efficiency score (AE) with the mean and standard deviation (mean std\text{mean}_{\,\text{std}}) evaluated across multiple trials (incorporating an extra trial with a different seed to ensure robustness) on MATH500, AIME 24--26, and AMC23. Average reasoning length (Len.) is reported as the mean token count rounded to the nearest integer.

Notation: M500 = MATH500; AIME = AIME 24–26; Avg = average across benchmarks. Fixed 1.25/1.50 = Ours (ℓ⋆\ell^{\star}) with fixed pressure 1.25/1.50.

MethodM500 Acc↑M500 Len↓M500 AE↑AIME Acc↑AIME Len↓AIME AE↑AMC23 Acc↑AMC23 Len↓AMC23 AE↑Avg Acc↑Avg Len↓Avg AE↑
Base76.29,70828.936,10280.020,94461.722,251
Fixed 1.2585.40.00\mathbf{85.4}_{0.00}2,0771.130.01\mathbf{1.13}_{0.01}31.7‾0.01\underline{31.7}_{0.01}8,6051.05‾0.02\underline{1.05}_{0.02}81.30.0181.3_{0.01}3,9580.840.020.84_{0.02}66.10.0166.1_{0.01}4,8801.02‾0.02\underline{1.02}_{0.02}
Fixed 1.5085.2‾0.01\underline{85.2}_{0.01}2,088‾\underline{2{,}088}1.130.14\mathbf{1.13}_{0.14}26.70.0126.7_{ 0.01}7,9570.370.190.37_{0.19}87.50.03\mathbf{87.5}_{0.03}3,5661.100.05\mathbf{1.10}_{0.05}66.5‾0.00\underline{66.5}_{0.00}4,5370.900.030.90_{0.03}
Ours (ℓ⋆\ell^{\star}+iso)84.80.0184.8_{0.01}2,2051.080.24‾\underline{1.08_{0.24}}35.00.02\mathbf{35.0}_{0.02}8,094‾\underline{8{,}094}1.390.19\mathbf{1.39}_{0.19}82.5‾0.03\underline{82.5}_{0.03}3,848‾\underline{3{,}848}0.89‾0.01\underline{0.89}_{0.01}67.40.01\mathbf{67.4}_{0.01}4,715‾\underline{4{,}715}1.150.11\mathbf{1.15}_{0.11}

To address this question, we conducted controlled experiments on DeepSeek-R1-Distill-Qwen-1.5B (150 steps) across multiple seeds and compared our method against both modest stronger fixed pressure (1.25, 1.50).

Intuitively, a higher, moderate fixed pressure (1.25, 1.50) might be expected to shorten reasoning lengths. Although the fixed gains of 1.25 and 1.50 exceed the maximum observed iso-pressure of 1.166 in Deepseek-1.5b, they do not guarantee shorter final reasoning length. Even with the same threshold-selection rule, pressure can indirectly affect subsequent threshold estimates through policy updates (Eq.16, Eq.12, Eq.5). The selected threshold need not decrease with increasing gain, so final reasoning length is not guaranteed to decrease monotonically.

And to evaluate the optimization stability of the proposed method, we ran additional trials under an independent random seed. Across seeds, Math-500 accuracy remains highly concentrated for all. In case of AMC23, 1.50 fixed pressure achieves the highest ae score. While a high fixed pressure (1.50) can occasionally yield strong performance on moderately challenging tasks like AMC23, such static configurations fail to generalize to more complex reasoning domains. Indeed, as reasoning complexity increases on AIME, the advantage of our dynamic calibration becomes clear. In contrast, the fixed pressure settings suffer from a severe performance drop on this benchmark, as clearly reflected in their significantly lower AE scores.

This consistent trend suggests that our dynamic thresholding and iso-pressure restoration provide a robust learning signal that effectively prevents accuracy degradation while reducing length.

Question 3. Additional Comparison with Related Methods.

  1. We have retrained QLPO, RoRecomp, and ARLCP on DeepSeek-R1-Distill-Qwen-1.5B using the same training and evaluation settings as OASIS. The results are reported in the table in common response. As detailed in our common response above, the results demonstrate that OASIS remains more effective than these closely related methods.

  2. SmartThinker's penalty strength is coupled to its calibration. Varying it independently would modify the original formulation and could violate its intended advantage-sign condition. We therefore retain its original calibration for the baseline comparison. Table 3 in the paper shows the contributions of the threshold candidates selection and iso-pressure restoration. The fixed-pressure ablation in our preceding response further supports adaptive pressure. With the same threshold-selection rule, iso-pressure achieves higher average accuracy and AE than the fixed gains.

Thanks for clarifying the scoring rule and adding the baseline comparisons. I’m raising my rating to 3. Please revise Algorithm 1 in the manuscript to show that all candidates use the same scoring reference.

Dear Reviewer,

Thank you again for your careful review and constructive suggestions. We have incorporated the revisions discussed in our response into the updated paper. Your thoughtful feedback has helped us improve the paper, and we sincerely appreciate the time and effort you devoted to reviewing our work.

Reviewer_2EjNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper addresses length-aware reinforcement learning for reasoning models, where a length-penalty threshold controls reasoning length. It proposes OASIS, a failure-conditioned adaptive thresholding method that uses both correct and incorrect rollout length distributions to select the penalty threshold, together with candidate threshold generation, a shared mixture objective with a local curvature filter, and batch-level iso-pressure restoration. It trains four open-source reasoning models with GRPO on DeepScaleR and evaluates on several popular math benchmarks.

Critical strengths
  1. The efficiency problem the article addresses is an important topic in the field of LLM reasoning and has strong impact.
  2. The authors provide complex mathematical derivations that support the reward design at the theoretical level, which I appreciate very much. 3. The article devotes considerable space to describing how to translate phenomena observed from data into algorithmic improvements at the theoretical level. The logic is coherent and rigorous, and the data-driven perspective has practical value.
  3. How to set the penalty threshold in length-aware RL is a good entry point for improving the algorithm at the reward level, and this research question is very interesting.
Critical weaknesses
  1. The experiments are limited to mathematical reasoning tasks. For other reasoning tasks, especially those whose reasoning structure is more open-ended and where verification of model trajectories is not binary as in mathematics, how should this method be transferred? Can the authors provide some ideas on this issue?
  2. The maximum generation length is 8192, whereas common practice for reasoning models is generally no less than 32768, which affects evaluation because there are more truncated trajectories. In fact, the results of the base models in the paper are indeed somewhat worse than those in the DeepSeek-R1 technical report (DeepSeek-R1-Distill series). Have the authors considered the impact of this on the experimental analysis?
  3. The algorithm uses a relatively large number of hyperparameters. Are there any empirical conclusions regarding the tuning of these hyperparameters?
Questions for authors

Please see Weaknesses. I have only minor concerns and look forward to the authors' response.

4 comments

We thank for your careful review and constructive feedback. We address each concern below.

Weakness 1: How this method can be transferred to other tasks

While our current evaluation focuses on mathematical reasoning, OASIS is fundamentally agnostic to the task domain and can be applied to any reasoning task where a correctness metric (even soft or model-based) is available.

Verifiable Reasoning Tasks (with Automated Evaluators): For tasks where solutions can be programmatically verified (e.g., formal logic or code generation), OASIS can be applied directly. Correctness is determined binarily via deterministic verification (e.g., unit tests), and length represents the intermediate reasoning trace. Since our method operates on the resulting success/failure signal rather than task-specific semantics, it can be applied to any verifiable domain.

Model-Based Verification (Reward Models & LLM-as-a-Judge): For open-ended reasoning (e.g., writing or planning), we can replace the binary rule with a classifying Reward Model or an LLM-as-a-Judge to partition rollouts. By thresholding the continuous reward scores (categorizing top-tier scores as ""Correct"" and lower-tier or flawed traces as ""Incorrect""), we can construct the same conditional length distributions, p(L∣C, x) and p(L∣I, x), to apply OASIS.

Weakness 2: Generation Length

First, we clarify that while the Appendix noted an approximate 100K limit, the evaluation was actually performed with a 120K maximum generation limit, providing more than sufficient headroom to ensure the reported metrics are free from evaluation truncation bias.

To address the concern about the 8K training ceiling, we conducted additional training under an expanded 32K length. Referring to the table below, the results confirm that our gains are not an artifact of tighter constraints. Even at 32K, OASIS continues to strike an effective balance between accuracy and conciseness. This indicates that our outcome-conditioned threshold mechanism identifies an intrinsically meaningful reasoning length, rather than relying on an external training cap to artificially truncate generation.

While comparing methods like SmartThinker also reduce reasoning lengths in the 32K setting, OASIS consistently delivers better reasoning accuracy with comparable or tighter length constraints. This demonstrates that our core mechanisms (candidate selection and iso-pressure restoration) remain robustly effective under expanded training regimes, further solidifying OASIS's distinct advantage in overall efficiency (AE Score).

Table: Comparison of accuracy and reasoning efficiency with 32K length training using DeepSeek-R1-Distill-Qwen-1.5B as the base model.

Notation: M500 = MATH500; AIME = AIME 24–26; Avg = average across benchmarks.

MethodM500 Acc. ↑M500 Len. ↓M500 AE ↑AIME Acc. ↑AIME Len. ↓AIME AE ↑AMC23 Acc. ↑AMC23 Len. ↓AMC23 AE ↑Average Acc. ↑Average Len. ↓Average AE ↑
GRPO81.410,231-31.135,674-77.514,596-63.320,167-
SmartThinker83.82,7930.98531.111,7080.88285.05,1870.895‾\underline{0.895}66.66,5630.940‾\underline{0.940}
Ours (ℓ⋆\ell^{\star})86.2‾\underline{86.2}2,816‾\underline{2{,}816}1.077‾\underline{1.077}35.6‾\underline{35.6}10,4721.388‾\underline{1.388}80.0‾\underline{80.0}6,4210.72567.3‾\underline{67.3}6,570‾\underline{6{,}570}0.901
Ours (ℓ⋆\ell^{\star}+iso)88.23,0591.14036.710,861‾\underline{10{,}861}1.47985.06,270‾\underline{6{,}270}0.92170.06,7301.030

Weakness 3: The number of hyperparameters and tuning tips.

We clarify that OASIS requires tuning only two or three hyperparameters, as most quantities appearing in the algorithm are automatically computed from rollout statistics rather than being independently tuned. These rollout-derived quantities include the candidate thresholds, soft-partition scale, mixture weights, and batch-level restoration factor. Indeed, as shown in Table 5, the only explicit method settings are the base gain (g), restoration ceiling (g_max), and minimum incorrect-sample count (min_inc), where the base gain (g) essentially serves as a fixed reference baseline rather than an active tuning parameter.

Restoration ceiling. The restoration ceiling safeguards against excessive pressure amplication when iso-pressure restoration produces a large gain. We fixed (g_max=2) across all models. Empirically, this ceiling never became active during the reported training runs. The restoration factor had a mean of 1.003–1.025 across runs and a maximum of 1.166. Thus, the reported results were not driven by clipping at a carefully selected ceiling.

Minimum incorrect-sample count. We chose (min_inc=2), the smallest count permitting a nonzero empirical variance for incorrect-length estimation.

In terms of parameter tuning, increasing min_inc is a possible adjustment that yields a more statistically stable estimation, though it may reduce the likelihood of selecting directional candidates, which is concurrently mitigated by increasing the rollouts.

Thanks for the authors’ response. In my view, a good method should minimize the use of hyperparameters as much as possible (since good performance should not rely on finely tuning hyperparameters). From this perspective, I still consider this a drawback.

However, the authors’ response has largely addressed my main concern, so I have decided to adjust my evaluation to reflect this.

We sincerely appreciate your reconsideration of our work and your decision to raise the rating.

We completely share the view that minimizing tuning-heavy parameters is highly desirable for practical deployment. When applying the OASIS mechanism, practitioners can comfortably use our default settings out-of-the-box, and if they wish to adjust anything for better statistical stability under different rollout sizes, min_inc is essentially the only parameter they would need to look at.

Thank you again for your time and valuable feedback, which helped us clarify this practical aspect of our method.

Reviewer_NoXnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces OASIS, a training method designed to make reasoning-focused language models think more efficiently without sacrificing accuracy. The main idea is that the model should learn not only from how long correct answers take, but also from how long incorrect answers take: if failures tend to be unusually long, the model should be encouraged to stop reasoning earlier, while if failures tend to be too short, it should be allowed more reasoning time. OASIS uses these patterns to automatically choose when a length penalty should begin during reinforcement learning, so unnecessary reasoning is discouraged without cutting off useful reasoning too soon. Experiments on four reasoning models and several math benchmarks, the method reduced reasoning length by about 32.7% on average while improving accuracy by 1.7 percentage points, suggesting that information from failed attempts can help models become both faster and more accurate.

Critical strengths

QLPO and RoRecomp already use the correlation between correctness and reasoning length, but they use it in a fixed, one-directional way: they assume the useful pattern is essentially "short correct responses are good; long incorrect responses are bad", and use that to resample or reprioritize training examples. OASIS instead treats the failure-length pattern as a signal for where to move the penalty threshold, and in which direction. If incorrect solutions tend to be longer than correct ones, OASIS moves the threshold down so the model is penalized earlier. But if incorrect solutions tend to be shorter, it moves the threshold up, allowing more unpenalized reasoning.

Critical weaknesses

The paper’s core novelty is using failure-length structure to bidirectionally relocate the reasoning-length penalty threshold. But it is not directly tested against the closest prior methods. In particular, QLPO, RoRecomp, and ARLCP are not included as matched baselines, even though they also exploit correctness–length relationships. This omission is impactful because the current experiments do not establish whether the proposed threshold-selection mechanism is actually superior to existing ways of using essentially the same underlying signal. A controlled comparison against these methods would strengthen the paper’s core novelty claim.

The method relies on an insufficiently validated interpretation of failure length. OASIS treats relatively short failures as evidence of underthinking and relatively long failures as evidence of overthinking, and uses this interpretation to move the reasoning-length threshold upward or downward. However, rollout length alone does not identify the cause of failure: a short failure may be a confident but incorrect solution, while a long failure may reflect intrinsic problem difficulty, an early strategic error, or productive but ultimately unsuccessful search. The paper does not provide intervention-based evidence or fine-grained per-problem analysis showing that increasing the threshold for short failures, or decreasing it for long failures, actually corrects the hypothesized failure mode.

Questions for authors

The advertised directional mechanism is rarely active on the Qwen models. In Table 6, the short- and long-failure directional candidates are selected in only 0.3% of all groups for Qwen3-1.7B and 0.1% for Qwen3-4B; even the whole incorrect-conditioned candidate is selected in only 3.5% and 0.9% of groups, respectively. Thus, most threshold decisions on these models revert to the marginal estimator. This weakens the attribution of gains to the paper’s central “unification” of short- and long-failure evidence, particularly because the directional-candidate ablation does not consistently improve performance across models.

1 comment

We thank for your careful review and constructive feedback. We address each concern below.


Weakness 1. Additional Comparison with Related Methods

We agree that direct comparisons are important for assessing the empirical value of our threshold-selection mechanism. We have therefore additionally trained QLPO, RoRecomp, and ARLCP on DeepSeek-R1-Distill-Qwen-1.5B using the same training and evaluation settings as OASIS. The results are reported in the table in our common response above. The results demonstrate that OASIS remains more effective than these closely related methods.

Weakness 2: Failure-Length Interpretation and Threshold Adaptation

We agree that the length of an individual failed rollout cannot by itself diagnose the underlying cause of failure (e.g., whether a short failure is a confident mistake or a long failure reflects intrinsic problem difficulty), which naturally makes per-problem causal diagnosis or intervention highly unreliable.

To clarify, OASIS is specifically designed to bypass the need for such unreliable per-problem causal diagnoses. Underthinking and overthinking in our paper do not just refer to rollout-level cognitive causes. Rather, they are simply statistical signals (empirical cases) observed when the incorrect length distribution is heavily concentrated on one side of the correct distribution.

Beyond these specific concentrated cases, OASIS is designed to mathematically encompass the entire spectrum of empirical layouts, including highly complex patterns like mixed failures (co-occurring short errors and long searches) or dispersed overlaps.

Question 1. Directional Candidate Selection and Its Contribution

OASIS applies the same candidate-generation and selection criteria across models, with eligibility and selection determined by each rollout group's outcome-conditioned length statistics. The unified mechanism allows whole-incorrect and directional corrections when supported by the evidence, while retaining the marginal estimator otherwise. Frequent directional selection across all models is therefore not a prerequisite of the design.

Table 3 in the paper shows that low selection frequency does not preclude empirical gains. Adding the whole-incorrect candidate improves both accuracy and AE on both Qwen models despite its low selection frequency. Adding the directional pair further improves Qwen3-1.7B by 0.5 pp in accuracy and 0.066 in AE. Although this incremental benifit is absent on Qwen3-4B, OASIS (ℓ⋆\ell^{\star}+iso) improves its accuracy from 91.9% to 93.0% and AE from 0.326 to 0.356 over the marginal-only baseline. Especially on DeepSeek-1.5B, where directional selection is more frequent, adding the directional pair produces a larger gain: 4.0 pp in accuracy and 0.207 in AE. These results support the usefulness of failure-length evidence for threshold selection. OASIS's unified framework enables this evidence to be used across models through a common selection rule, without requiring a model-specific choice of correction strategy.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.