acceptodds
Under review as a conference paper at ICLR 2027

Understanding Enrichment in Reinforcement Learning

Abstract

When rewards are sparse, reinforcement learning with verifiable rewards (RLVR) often uses hints or intermediate guidance to generate more successful rollouts. This enrichment biases policy-gradient updates unless corrected via importance weights, but existing methods omit correction or truncate importance weights in order to avoid the high variance of correction. It thus remains unclear what exactly is gained or lost in RLVR by correcting enriched rollouts. We show, mathematically, that omitted or truncated correction implicitly reweights the defined reward, and we decompose the resulting gradient error into scale, rotation, and variance to explain their distinct effects on learning. To make correction practical, we develop a novel sequential Monte Carlo (SMC) weight correction mechanism that, under stability assumptions, tempers the exponential compounding of standard correction variance over the length of a sample to an additive accumulation. We then apply our analysis of enrichment to interpreting the results of a fine-tune of Qwen3-1.7B on a sparse band of OpenMathReasoning, establishing a concrete mechanism of how enrichment, both corrected and uncorrected, helps avoid collapse in sparse domains. Our main contribution is to understand, in general, how enrichment and correction can affect training, as opposed to claiming that either mode of operation is superior to unenriched RL.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning with Verified Rewards, Sparse Domains, Weight Correction, SMC Sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fcGoOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how enriched sampling, such as generating responses with hints, changes reinforcement learning with verifiable rewards. It interprets uncorrected enrichment as implicit reward reweighting and decomposes gradient error into scale, rotation, and variance. The authors propose an SMC-based correction mechanism with mean-preservation guarantees for the analyzed REINFORCE estimator and variance bounds under additional assumptions. Controlled experiments illustrate these effects, followed by a Qwen3-1.7B case study on a low-success-rate subset of OpenMathReasoning investigating enrichment and training collapse.

Critical strengths
  1. Useful conceptual clarification. The reward-reweighting interpretation and the separation of scale, rotation, and variance provide an accessible framework for understanding enriched training. In particular, the analysis clarifies how enrichment can change the relative strength of reward and KL regularization, or preferentially reinforce particular solution families.

  2. Carefully designed controlled experiments. The experiments isolate several sources of gradient distortion. Appendix D.2 holds reward sparsity fixed while varying sequence length and demonstrates substantial variance reduction from SMC in the constructed setting. The analysis also explicitly accounts for shared ancestry and the qualifications needed for the variance bound.

Critical weaknesses
  1. Limited empirical support for the robustness of the collapse-prevention explanation. The sole real-model experiment uses Qwen3-1.7B on a hard, low-success-rate mathematics subset, with 500 training and 300 evaluation problems. I recognize that Section 4 is positioned as a case study. Nevertheless, the experiment leaves unclear which aspects of the observed behavior depend on the initial success rate, guidance strength, or model capacity. The SMC arm’s pass@1 advantage over unenriched training is also uncertain: the reported 95% bootstrap interval is [−0.21, +1.12] percentage points, with a one-sided p-value of 0.09. Although the authors report significant pass@k gains, the corresponding statistical details and seed-level variability are insufficiently specified. Together, these limitations constrain the empirical support for applying the proposed explanation beyond this particular setting.

  2. The proposed anti-collapse mechanism is not directly established by the reported measurements. Section 4 attributes the results to increased policy entropy and learning new successes. However, the description of Figure 4(c) does not specify the sampling and scoring distributions or whether correction weights are applied. Unless the statistic is computed on separately generated unguided base-policy rollouts or with an appropriate correction, changes in the training-sample distribution can also affect it. Consequently, decreasing typicality does not by itself establish increased base-policy entropy. Similarly, pass@k trends are consistent with changes in success coverage but do not uniquely identify memorization of known solutions. Since Section 4 uses these observations to explain collapse prevention, clarifying the measurement protocol and providing direct entropy and coverage diagnostics would strengthen this central claim.

Questions for authors
  1. Please specify the sampling and scoring distributions used in Figure 4(c), whether SMC/importance weights are applied to this statistic, and whether it is evaluated on fresh unguided base-policy rollouts.

  2. For which values of k are the pass@k improvements statistically significant? Please specify the confidence intervals, bootstrap resampling unit, and treatment of the three training seeds.

  3. Given the qualification in Section 3.1 that mean preservation does not automatically extend to GRPO, what is the exact SMC-corrected GRPO update used in Section 4? Please clarify how correction weights enter the group mean, standard deviation, and final loss, and how this update relates to the cited weighted-surrogate interpretation.

  4. Could existing checkpoints be used to measure base-policy entropy and per-problem success coverage under matched prompts and unguided decoding settings? If feasible, a targeted analysis stratified by initial success rate would help assess whether the collapse-prevention explanation consistently matches the observed training behavior.

3 comments

We thank the reviewer for their review.

Limited empirical support for the robustness of the collapse-prevention explanation…

The proposed anti-collapse mechanism is not directly established by the reported measurements… providing direct entropy and coverage diagnostics would strengthen this central claim.

  1. Please specify the sampling and scoring distributions used in Figure 4(c), whether SMC/importance weights are applied to this statistic, and whether it is evaluated on fresh unguided base-policy rollouts.

Figure 4(c) reports the mean length-normalized log probability under each run’s unhinted policy, using training responses drawn from the base policy for unenriched training, the enriched policy for uncorrected training, and the resampled terminal population for SMC. No SMC or importance weights are applied, and we did not generate fresh unguided evaluation rollouts. We have conducted an additional per-problem success coverage and entropy evaluation to substantiate. Please refer to tags [CVRG] and [ETRP] for details. We will revise the entropy and memoization interpretation to ground it concretely in our new coverage experiments.

  1. For which values of k are the pass@k improvements statistically significant? Please specify the confidence intervals, bootstrap resampling unit, and treatment of the three training seeds.

We provide the analysis below, where [Config - unenriched] indicates the difference in pass@k coverage, in percentage points, with 95% confidence intervals:

kSMC − unenrichedUncorrected − unenriched
1+0.46 [−0.20, +1.13]+0.75 [+0.18, +1.30]
2+0.93 [−0.16, +2.04]+1.56 [+0.62, +2.48]
4+1.83 [+0.20, +3.50]+3.10 [+1.68, +4.53]
8+3.33 [+1.08, +5.63]+5.61 [+3.64, +7.62]
16+5.39 [+2.44, +8.39]+8.91 [+6.31, +11.58]
32+7.33 [+3.44, +11.22]+12.11 [+8.67, +15.56]

SMC surpasses unenriched with statistical significance at all tested k ≥ 4, while uncorrected surpasses unenriched across all tested k. Pass@k is computed separately for each problem and seed from evaluation responses following the method in Humaneval [2], then we average the three seed-specific values for each problem. Bootstrapping follows the method in [3] over 100000 runs that draw 300 problems with replacement over the full 300-problem eval set.

  1. Given the qualification in Section 3.1 that mean preservation does not automatically extend to GRPO, what is the exact SMC-corrected GRPO update used in Section 4? Please clarify how correction weights enter the group mean, standard deviation, and final loss, and how this update relates to the cited weighted-surrogate interpretation.

Please refer to Reviewer [k8go]’s question on the GRPO update, tagged [GRPU]. The cited paper [1] derives a weighted group-relative update in Section 4.3, equation (9); our corrected reward update follows this form using normalized SMC correction weights, with an additional scaling term from the group reward’s standard deviation.

  1. Could existing checkpoints be used to measure base-policy entropy and per-problem success coverage under matched prompts and unguided decoding settings? If feasible, a targeted analysis stratified by initial success rate would help assess whether the collapse-prevention explanation consistently matches the observed training behavior.

Please refer to Reviewer [k8go]’s question on coverage, tagged [CVRG], for the requested per-problem success coverage experiment, which was performed on the base policy and all three trained policies (unenriched, uncorrected, and SMC).

[ETRP]

We also provide a comparison of entropy using the same experimental settings as in [CVRG] (using a set of fresh 64 responses sampled from each policy). Uncorrected and SMC training have statistically significant higher entropy compared to unenriched training:

ComparisonΔ entropy/token, 95% CIΔ response entropy, 95% CI
Uncorrected−unenriched+0.157 [0.145, 0.168]+186.32 [168.89, 203.74]
SMC−unenriched+0.219 [0.205, 0.233]+207.06 [190.97, 223.16]

The difference is statistically significant over both groups in [CVRG] (those where the initial policy produced a success and those where it did not). We also note that unenriched training decreases per-token entropy compared to the initial policy by 52% (from 0.299 nats to 0.144 nats), while both SMC (0.363 nats) and uncorrected training (0.300 nats) maintain the statistic.

We do agree that while the entropy difference is statistically significant, it does not establish a causation relationship with the increase in coverage. We will revise the claim such that the coverage explanation is directly supported by the experiment from [CVRG], and include the entropy comparison as a separate observation.

References

[1] https://arxiv.org/abs/2509.24203v2

[2] https://arxiv.org/abs/2107.03374

[3] https://aclanthology.org/D12-1091/

Thank you for your time and feedback. We have posted our responses to your concerns. As the discussion period ends soon, we would appreciate it if you could take a look and let us know if you have any further questions.

The new coverage and entropy results address my main concerns. I will raise my rating to 3.

Reviewer_e3nEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper first studies how enriched rollout sampling can help and improve the performance of RLVR. Then it proposed a new method SMC based correction mechanism to reduce the variance. The empirical study compares couple basedlines and do show enrichment can benefit training performance.

Critical strengths
  1. paper addresses an important issue and the reward-reweighting view gives a clean explanation for why uncorrected enriched rollouts can help, hurt, or optimize a different effective objective.

  2. The SMC construction is interesting and directly targets multiplicative importance-weight accumulation.

Critical weaknesses
  1. The theoretical analysiss is intuitive but the example are too simplified to reflect actual training. The controlled examples isolate each component using short tabular policies, exact gradients, normalized update lengths, and deliberately constructed enrichment, but do not study their interaction under shared neural parameters, adaptive optimization, or rollout policies. The paper should discuss the interactions among these effects and validate its key assumptions on actual rollouts.

  2. The paper did show enrichment is useful and can benefit training. However, the proposed SMC correction have very similar performance compare with the uncorrected method. I do suggest auther can provide some statistical analysis or result to show the significance.

Questions for authors
  1. For Figure 1, the original objective does not have α\alpha term in the objective. So it is initially unclear how can the value vary from different α\alpha? What my iterpertation is the θ\theta should be dependent on α\alpha? I think auther may better work on the notation to avoid confusion.

  2. Figure 4 can support "Enrichment can be helpful for real training, especially in very sparse domains," However, it did not show "enrichment provides access to new successes the base policy would have failed to generate other- wise." Since decreasing in training-sample typicality does not necessarily demonstrate that the enriched policy learns successes that the base policy cannot generate.Can auther provide more details how they can get this conclusion from Figure 4?

  3. Do the key assumptions behind the SMC variance bound, particularly bounded importance ratios, accurate future-reward estimates, and favorable particle ordering, hold on actual LLM rollouts? what are their limitation?

2 comments

We thank the reviewer for their review.

The theoretical analysiss is intuitive but the example are too simplified to reflect actual training... The paper should discuss the interactions among these effects and validate its key assumptions on actual rollouts.

Our paper aims to establish a theory on how enrichment affects RL, where the goal of the controlled experiments are to show how the theoretical predictions manifest in practice without training dynamics obscuring the analysis. We observe that an experiment with real training dynamics may have the opposite problem: a difference in the end result may be attributed to either training dynamics or enrichment, preventing us from providing a concrete evidence of how enrichment impacts training. Our theory shows that the scale and rotation factor change the effective reward being optimized for, and that the rotation and variance factors can significantly change the trajectory by changing the direction of local updates and the effect of stochasticity. Section 4 illustrates how this prediction manifests in practice over a real training run, showing that uncorrected / SMC-corrected enrichment result in significantly different results compared to unenriched training.

The paper did show enrichment is useful and can benefit training... I do suggest auther can provide some statistical analysis or result to show the significance.

The goal of our SMC correction is to propose a correction mechanism that tempers the variance cost of correcting enriched rollouts, which may be useful in the case one wishes to perform correction (i.e.., wishes to recover the base-policy REINFORCE gradient) . Our paper does not establish, nor suggest, that uncorrected enrichment is inferior compared to corrected enrichment or unenriched training.. The analysis in Section 2 shows that uncorrected enrichment is training over a different objective, which may become beneficial for sparse-reward domains through the coverage mechanism illustrated in Section 4.

  1. For Figure 1, the original objective does not have term in the objective. So it is initially unclear how can the value vary from different ? What my iterpertation is the should be dependent on ? I think auther may better work on the notation to avoid confusion.

The reviewer is correct: the parameters θk\theta_k depend on the scale coefficient α\alpha. We will update the notation to θkα\theta_k^{\alpha}.

  1. Figure 4 can support "Enrichment can be helpful for real training, especially in very sparse domains," However, it did not show "enrichment provides access to new successes the base policy would have failed to generate other- wise." Since decreasing in training-sample typicality does not necessarily demonstrate that the enriched policy learns successes that the base policy cannot generate.Can auther provide more details how they can get this conclusion from Figure 4?

We have conducted an additional per-problem success coverage evaluation to substantiate the claim. Please refer to the response for Reviewer [k8go]’s question on coverage, tagged [CVRG]. We will update the paper to better substantiate the claims.

  1. Do the key assumptions behind the SMC variance bound, particularly bounded importance ratios, accurate future-reward estimates, and favorable particle ordering, hold on actual LLM rollouts? what are their limitation?

Please refer to Reviewer [k8go]’s question on SMC variance, tagged [VRNC]. Our variance bound does not require accurate future-reward estimates. The reviewer is correct that not all of the variance bound assumptions hold on actual LLM rollouts; but the role of the variance theorem is to predict and quantitatively justify that SMC resampling can bring a significant reduction in variance. We will update our claims to make the distinction explicit.

Thank you for your time and feedback. We have posted our responses to your concerns. As the discussion period ends soon, we would appreciate it if you could take a look and let us know if you have any further questions.

Reviewer_k8goOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper looks at what happens when we generate RL rollouts with hints or other guidance, then use them to train the unguided policy. The authors call this enrichment. They explain how skipping importance correction changes the update, split the effects into scale, rotation, and variance, and propose SMC correction that resamples partial responses. The corrected REINFORCE estimator keeps the right mean. Small controlled experiments illustrate the different effects, and a Qwen3-1.7B experiment on hard math problems compares ordinary training with corrected and uncorrected enrichment, including what happens to solution coverage.

Critical strengths
  • This is a relevant question. Guided rollouts are useful.
  • The paper doesn't assume unbiased is always better. It explains both why uncorrected enrichment can help and why exact correction can be unstable.
  • The SMC analysis takes shared ancestry and resampling noise seriously.
Critical weaknesses
  • I'm still unsure how much is new in the basic analysis. The effective-reward rewrite and the parallel/orthogonal/variance split are fairly standard. Putting them together is useful, but I'd like a clearer statement of what we can now predict or understand that we couldn't before. I'm not asking this paper to beat SOTA; I do want more from the central explanation.
  • The additive-in-length variance message needs more qualification up front. The full bound still has a factor exponential in length divided by particle count, and the score bound can itself grow with length. There is also a particle-order condition involving future expected gradient contributions. The appendix explains these limits, but I came away from the main text with a stronger impression. At fixed particle count, exponential horizon dependence hasn't been removed in general.
  • I don't see the full connection to the GRPO experiment yet. The mean-preservation result is for REINFORCE. Group normalization and correlations between SMC samples add other issues, which the authors acknowledge. But I still need the exact loss used in the LLM run, including where the correction weights enter, to tell what is being corrected there.
  • The LLM result is fairly limited: one model and one small hard subset. The corrected method's pass@1 improvement over unenriched training isn't significant at the reported 95% interval. The pass@k result is worth showing, but typicality and pass@k alone don't establish the proposed coverage explanation. Per-problem success coverage and solution-diversity measurements would make that argument much stronger.
  • Comparing the same number of updates doesn't settle the cost question. Hints, probability evaluation under both policies, and resampling all cost something. I'd like token counts, forward-pass counts, memory use, and wall-clock time before judging whether full correction is worth it in practice.
Questions for authors
  1. What are the particle count, ESS threshold, resampling schedule, and enrichment potential in the LLM experiment? How are the hints made?
  2. What exactly is the corrected GRPO loss? Which parts of the REINFORCE guarantee apply to it?
  3. What do the actual weight distributions, ancestry diversity, and resampling frequency look like on LLM responses? Is there any evidence that the particle-order condition is reasonable there?
  4. How does this compare with truncated correction or success-filtered SFT on the same guided samples, with the same total generation budget?
  5. Does the coverage story still hold if entropy or KL is matched? Can you show coverage per problem rather than only aggregate pass@k?
3 comments

We thank the reviewer for their review.

I'm still unsure how much is new in the basic analysis. … I'm not asking this paper to beat SOTA; I do want more from the central explanation.

While it is true that the decomposition of the update gradient into scale, rotation, and variance is a standard mathematical technique, we would like to point out that we are the first to connect this three-point decomposition to study enrichment in RL. Our analyses predict that:

  • Given a scale factor α\alpha, enrichment penalizes the regularization strength by 1α\frac{1}{\alpha};
  • Rotation identifies exactly how the training trajectory can remain different even under a perfect trainer with no estimation noise,
  • The variance analysis derives how variance correction can leave finite training batches unrepresentative of the true corrected gradients.

Overall, the analysis establishes how simple intervention techniques fail to “correct” enriched training, and thus that uncorrected enriched training should be better accepted as a separate training mechanism rather than one following standard RL intuition (i.e., maximizing the target reward). We propose the SMC corrector as one way that enriched training can regain this intuition. We do NOT claim that unenriched or SMC training is better than uncorrected training: the experiments from Section 4 illustrate that despite losing the intuition of standard RL, uncorrected training can achieve better performance on certain tasks, and we provide an analysis of why that may be the case.

The additive-in-length variance message needs more qualification up front. … At fixed particle count, exponential horizon dependence hasn't been removed in general. 3. What do the actual weight distributions, ancestry diversity, and resampling frequency look like on LLM responses? Is there any evidence that the particle-order condition is reasonable there?

[VRNC]

The weight distribution is:

PercentileWeight
10th0.624
Median1.022
90th1.236
99th2.034
Maximum2.990

The largest weight in a group accounts for 9.9% of its total weight on average, illustrating that the SMC corrector achieves the intended effect of reducing weight variance. An average of 2.13/16 ancestors survive in during SMC resampling, and resampling occurred at 55.1% of possible locations (an average of 15.19 resamples / 27.55 opportunities).

We agree that the additive-in-length variance claim depends on assumptions that do not always hold in practice. The main role of the theorem is to predict and quantitatively justify that SMC resampling can bring a significant reduction in variance. The comparison in appendix D.2 illustrates this fact, with SMC achieving a 3929x reduction in variance. We will update the paper to clearly state the limitations and the role of the additive theorem.

I don't see the full connection to the GRPO experiment yet… But I still need the exact loss used in the LLM run, including where the correction weights enter, to tell what is being corrected there.

  1. What are the particle count, ESS threshold, resampling schedule, and enrichment potential in the LLM experiment? How are the hints made?
  1. What exactly is the corrected GRPO loss? Which parts of the REINFORCE guarantee apply to it?

[GRPU]

Particle count is 16 per problem, the ESS threshold is 12.8 (80% of the particle count), resampling opportunities occur every 64 tokens. Hints supply the first half of a known verified solution as additional prompt context. Enrichment occurs by evaluating the target model with and without the hint, then blending their next-token logits in a 3:5 ratio before sampling. The actual GRPO loss for the experiment in Section 4 follows the loss defined in Appendix A, replacing its advantage coefficient Âᵢ by

{\sqrt{\frac{N}{N-1}\mu_\omega(1-\mu_\omega)}+10^{-6}}, \qquad \omega_i=\frac{Nw_i}{\sum_jw_j}, \qquad \mu_\omega=\frac1N\sum_j\omega_jR(x_j),$$ Where w_i is the terminal SMC correction weight. At all w_i = 1, the advantage coefficient reduces to that in Appendix A. We also add a reference-KL penalty with coefficient 0.001. This loss is averaged over response tokens. Section 3 explains the mean-preservation guarantee to REINFORCE and discusses the justification for applying this correction to GRPO. We will update the paper to include the actual implementation loss as well. **This response is continued on the next comment.**

The LLM result is fairly limited… Per-problem success coverage and solution-diversity measurements would make that argument much stronger.

  1. Does the coverage story still hold if entropy or KL is matched? Can you show coverage per problem rather than only aggregate pass@k?

[CVRG]

We provide an additional per-problem success coverage experiment over 4 policies: the base policy, and the three trained policies (unenriched, uncorrected, and SMC) from Section 4. For each entry from our 300-problem held-out evaluation set, we sampled 64 fresh responses, then computed the fraction of problems each policy solved at pass@32 using the method from Humaneval [1]. The 300 problems are further broken down into 2 groups: depending on whether the initial policy solves a problem within 32 attempts (86 problems) or not (214 problems), following the initial success rate stratification requested by [fcGo].

PolicyAll problemsNo initial observed success (214)At least one initial success (86)
Initial26.69%7.75%73.84%
Unenriched24.75%11.60%57.46%
Uncorrected33.78%16.83%75.97%
SMC31.25%14.32%73.35%

Both enriched methods improve overall coverage vs. unenriched training statistically significantly (the intervals are [2.48, 10.51] for SMC and [5.25,12.81] for uncorrected). In particular, the breakdown shows unenriched training decreases coverage on problems that the initial policy could solve, while SMC and uncorrected training retain coverage. In addition, uncorrected achieves a statistically significant increase in coverage rate on the first group compared to unenriched training. We will restate our coverage claim to be more precise with this table. Please also refer to reviewer [fcGo]’s question on entropy, tagged [ETRP], for a computation of entropy over the same fresh 64 responses. SMC and uncorrected training both significantly increase entropy compared to unenriched training. Our paper is centered on the effect of enrichment on learning, and we believe the above coverage experiment illustrates that enrichment improves coverage (on the setting in Section 4) clearly. We will revise the mention of entropy as the mechanism with the concrete table above.

Comparing the same number of updates doesn't settle the cost question…

We provide a comparison of GPU hours on A100s required for training, which is the most relevant cost metric:

UnenrichedUncorrectedSMC
28.0764.6261.14

Our token count logging is incomplete for SMC, so we provide upper and lower bounds of generated tokens (the real count of tokens generated by SMC is somewhere in between). For uncorrected and unenriched training, the token counts are exact.

UnenrichedUncorrectedSMC
107.57M96.60M39.15–98.53M

SMC adds nearly no overhead compared to uncorrected enrichment. (Both uncorrected and SMC add overhead vs. unenriched training due to the standard enrichment cost). Memory use was also roughly equivalent across all three methods. We did not have to use additional hardware to support the enriched configs.

  1. How does this compare with truncated correction or success-filtered SFT on the same guided samples, with the same total generation budget?

We ran a separate experiment using truncated correction, same three seeds, where token-level correction weights were capped at 2. |Truncated - config| indicates the difference between the new truncated run and a training configuration.

ComparisonΔ pass@1 (pp), 95% CIΔ pass@32 (pp), 95% CI
Truncated − IID+0.78 [+0.16, +1.39]+8.11 [+4.56, +11.67]
Truncated − uncorrected+0.18 [−0.24, +0.61]−3.89 [−6.78, −1.00]
Truncated − SMC+0.32 [−0.10, +0.75]+0.78 [−2.22, +3.78]

Like uncorrected enrichment, truncation displays an advantage over unenriched for both pass@1 and pass@32. Comparisons between truncated and SMC / uncorrected are statistically insignificant, except for the fact that uncorrected outperforms truncated on the pass@32 metric.

When update token counts are matched, the comparison between truncated and SMC / uncorrected becomes statistically insignificant:

Budget / comparison armTruncated pass@1Comparison pass@1Δ pass@1 (pp), 95% CITruncated pass@32Comparison pass@32Δ pass@32 (pp), 95% CI
Match uncorrected’s generated-token count3.21%3.19%+0.02 [−0.43, +0.47]34.67%36.11%−1.44 [−4.44, +1.56]
SMC upper-bound generated-token allowance3.28%3.05%+0.23 [−0.18, +0.64]33.00%31.44%+1.56 [−1.44, +4.67]

Thanks for the authors' response, my concerns are mainly addressed, I'll increase my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.