acceptodds
Under review as a conference paper at ICLR 2027

The Distillation Game: Adaptive Attacks & Efficient Defenses

Abstract

Distillation attacks create a deployment trade-off for model providers: the same outputs that make a model more useful can also make it easier to imitate. We study this trade-off through a minimax game between a utility-constrained teacher and an adaptive student. Our framework yields tractable one-sided response rules: an adaptive evaluation rule in which the student reweights high-value examples, and a teacher-side defense template that suppresses outputs most useful for distillation. From a cheap proxy for example value, we derive Product-of-Experts (PoE), a simple forward-pass-only defense that combines the teacher with a proxy student during generation. Empirically, adaptive evaluation reveals a large passive-adaptive gap: on state-of-the-art defenses, adaptive students recover substantially more capability than passive evaluation suggests on GSM8K and MATH. Under this stronger evaluation, the apparent robustness gap between expensive defenses and PoE narrows considerably, while PoE remains substantially cheaper and preserves higher-quality reasoning traces. Overall, our results suggest that strong distillation remains difficult to stop, and that progress on antidistillation should be judged against adaptive students rather than passive ones.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
distillation, antidistillation, LLM security, privacy

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_y1tZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies antidistillation through a game-theoretic framework between a utility-constrained teacher and an adaptive student. The teacher releases a defended output distribution while remaining close to a reference model, and the student adaptively reweights released examples under a KL-constrained adaptation budget. The framework yields exponential-tilting best responses for both sides. On the attacker side, the student emphasizes examples with high estimated downstream learning value; on the defender side, the teacher suppresses examples that are most useful for distillation. Based on a likelihood-gap proxy for example value, the authors derive a lightweight Product-of-Experts (PoE) defense that combines teacher and proxy-student probabilities during decoding.

Critical strengths
  1. The paper identifies an important evaluation weakness in prior antidistillation work by showing that adaptive students can recover substantially more capability than passive students.
  2. The minimax formulation provides a clean and interpretable framework that unifies adaptive sample reweighting and teacher-side defense under explicit fidelity and adaptation budgets.
  3. The proposed PoE defense is computationally attractive, requiring only forward passes while achieving adaptive leakage comparable to the substantially more expensive ADS baseline.
Critical weaknesses
  1. The paper does not sufficiently distinguish the one-sided teacher best response used to motivate PoE from the solution of the full Stackelberg minimax game.
  2. The statement of Theorem 3.1 appears to omit assumptions needed for claims such as uniqueness of the dual variables and for the finite-dimensional convex analysis used in the proof.
  3. The implemented adaptive attack differs from the population-level KL-constrained best response because it dynamically recomputes values and normalizes weights within minibatches, while the resulting effective adaptation budget is not reported.
  4. The empirical evidence is limited to two mathematical reasoning benchmarks and a single main teacher/student configuration, making the generality of the conclusions unclear.
Questions for authors
  1. What additional assumptions are needed for Theorem 3.1, especially for the uniqueness of the dual variables?
  2. Can you report the effective KL adaptation budget of the adaptive attack used in the experiments?
  3. How sensitive is the adaptive attack to the amount of held-out downstream data used to compute the value scores?
1 comment

We thank the reviewer for engaging, and respond to their comments below. We also refer the reviewer to the revised version of the submission we uploaded.

The paper does not sufficiently distinguish the one-sided teacher best response used to motivate PoE from the solution of the full Stackelberg minimax game.

We agree that this distinction should be clear, and we note that it is already stated immediately after Theorem 3.1, but we will recall it again when introducing PoE. For reference, recall that PoE instantiates the one-sided teacher response for a fixed proxy value function, and we do not claim it solves the full Stackelberg minimax game, which is analyzed separately in Appendix B.

What additional assumptions are needed for Theorem 3.1, especially for the uniqueness of the dual variables?

We agree that Theorem 3.1 could be clearer about the active KL constraints. We added a finite data space (i.e., the prompt-output pairs) assumption for simplicity, noting that general prompt distributions can be handled under suitable measurability and integrability conditions. We also dropped the parameter uniqueness claim, which is unnecessary for deriving the best-response formulas. Boundary cases such as zero or sufficiently large KL budgets are now treated separately in the appendix.

The implemented adaptive attack differs from the population-level KL-constrained best response [...]

We acknowledge in Section 3.1 that our attack is a practical approximation of the theoretical best student response, given that computing the exact population best response requires scoring and normalization over the teacher’s full output distribution, which is computationally impractical. We therefore normalize within minibatches and update values as the student learns. This suffices for our empirical goal of demonstrating that passive evaluation can underestimate leakage, and is informed by our theoretical framework.

The empirical evidence is limited to two mathematical reasoning benchmarks and a single main teacher/student configuration, making the generality of the conclusions unclear.

We agree that our empirical results focus on mathematical reasoning, which we acknowledge in the introduction. There, we mention that we follow Savani et al. (2025)’s mathematical reasoning setting to measure the influence of attacker adaptation only, keeping the rest fixed. Regarding teacher/student configurations, we additionally show in Appendix C.4 an adaptive–passive gap with a smaller defense proxy. Evidently, broader task and teacher/student coverage would be great future work.

[...] I would nevertheless encourage the authors to explicitly acknowledge the dual-use nature of the adaptive attack [...]

We agree and now acknowledge in the revised submission that the distillation attack could be used by strategic actors, although our purposes are scientific.

Can you report the effective KL adaptation budget of the adaptive attack used in the experiments?

The KL adaptation budget is mostly a theoretical instrument to model the fact that the adversary may adapt (or not) to different degrees. This budget is dual (i.e. its Lagrange multiplier) to the “sharpness” hyperparameter, for which we found 0.5 to be usually a good value.

How sensitive is the adaptive attack to the amount of held-out downstream data used to compute the value scores?

We ran a new ablation suggesting that substantially fewer holdout traces suffice in this setting. We varied the holdout size on GSM8K with ADS λ=0.05\lambda=0.05, β=0.5\beta=0.5, keeping the student initialization, training budget, and 1,209 test questions fixed.

Holdout tracesStudent accuracy
Full: 2,23552.36%
1,00052.52%
50051.20%
25054.18%
10053.10%

With 100 holdout traces, which is less than 2% the total data available to the student for SFT, we observe no significant accuracy drop. Note this is for a single seed so small differences are insignificant.

Reviewer_u79VOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies defenses against model distillation through a minimax game between a utility-constrained teacher and an adaptive student. The teacher chooses a released policy that stays close to a reference teacher, while the student is allowed to concentrate training mass on examples with high downstream learning value subject to a KL adaptation budget (Sec. 2). Theorem 3.1 derives exponential-tilt one-sided best responses for both players. On the attacker side, this motivates an adaptive distillation procedure based on gradient-alignment scores; on the defender side, substituting a teacher-versus-proxy likelihood-gap value function yields a forward-pass-only Product-of-Experts (PoE) decoding rule (Sec. 3). Empirically, the paper evaluates GSM8K and MATH with a DeepSeek-R1-Distill-Qwen-7B teacher and a Llama-3.2-3B distilled student. Adaptive weighting substantially increases recovered student capability relative to passive distillation, especially for defended teachers (Table 1; Fig. 3). PoE is competitive with ADS under adaptive evaluation while using less generation time, and it preserves more auditable reasoning traces according to the paper's trace-quality evaluation (Fig. 4). Appendix C further includes filtering, proxy-size, frontier-model distillation, and trace-rewriting analyses.

Critical strengths
  • S1: The paper identifies an important evaluation weakness in current antidistillation work: a student that passively trains on all released traces is not necessarily the strongest attacker permitted by the release. The empirical effect is large enough to matter qualitatively. For example, at the representative GSM8K point in Table 1, ADS student accuracy rises from 34.33% under passive training to 51.50% under adaptive training; the same direction is visible for PoE and on MATH. This makes the recommendation to evaluate defenses against adaptive students practically consequential rather than merely conceptual.
  • S2: The minimax formulation provides a clean common language for attacker adaptation and teacher-side defense. In particular, Theorem 3.1 gives exponential-tilt best responses under the KL-constrained formulation, which yields a direct adaptive evaluation rule on the student side and gives a useful interpretation of gradient-based defenses such as ADS on the teacher side. Appendix A supplies the one-sided proofs, and Appendix B extends the analysis to existence/characterization of an optimal pair in the finite setting. This theoretical organization is useful even beyond the specific PoE instantiation.
  • S3: PoE is a simple and practically motivated defense derived from the same framework. Replacing the expensive gradient-based value with the teacher-versus-proxy log-likelihood gap leads to the geometric mixture in Eq. 12 and a forward-pass-only decoding rule (Algorithm 2). The resulting trade-off is interesting: at the representative GSM8K operating point, PoE has similar adaptive leakage to ADS (49.46% vs. 51.50% student accuracy) while the reported generation-time overhead is much smaller (1.64x vs. 2.93x; Table 1). The paper appropriately avoids claiming uniform dominance.
  • S4: The evaluation goes beyond a single leakage number and explicitly considers teacher utility, adaptive leakage, generation cost, and released-trace auditability. This is a stronger evaluation perspective than judging a defense only by passive student accuracy. The trace-quality analysis in Fig. 4 is also accompanied by a blinded human sanity check of the automated rubric, reporting quadratic-weighted kappa of 0.76 on 30 traces, which is a useful validation step.
Critical weaknesses
  • W1: My technical concern is the connection between the population game and the implemented adaptive attack. In Eq. 3 and Eq. 5, the prompt marginal is fixed as x ~ D and the attacker changes only the conditional output distribution pi_eff(y|x). Consequently, Eq. 6 is normalized separately for each x. Algorithm 1, however, assigns a single normalized weight across all (x, y) examples in a minibatch, and Sec. 4.1 states that the release contains a reasoning trace for each prompt. With one sampled trace per prompt, this minibatch weighting appears to change the effective prompt distribution as well as the output distribution, and it does not estimate the per-prompt normalizer in Eq. 6. As written, I am therefore not fully convinced that the empirical attack is literally a stochastic implementation of the stated best response. A joint-distribution formulation over (x, y), or a clearer derivation connecting Algorithm 1 to the conditional game, would resolve this issue.
  • W2: The information available to the adaptive attacker is underspecified, and this matters for interpreting the passive-adaptive gap. Algorithm 1 requires a downstream objective L(theta), with a held-out validation NLL given as the example, and repeatedly uses its gradient to score released traces. Appendix C.1 defines sizable train/holdout/test splits and gives the adaptive sharpness parameter, but I could not find a precise description of what L is in the reported experiments, whether ground-truth holdout answers are used, how often this gradient is recomputed, or how this extra supervision/computation is accounted for in the threat model. If labeled holdout data are used, the adaptive student has task supervision that the passive baseline does not explicitly exploit. This does not invalidate the attack, but it is important for attributing the gain specifically to adaptive reweighting and for reproducing the result.
  • W3: The empirical evidence for the broad methodological conclusion is still fairly concentrated. The main adaptive-defense experiments use two closely related mathematical-reasoning benchmarks and one principal teacher/proxy/final-student stack (Sec. 4.1). Appendix C.4 varies the proxy size but keeps the same teacher and final student, while Appendix C.2 shows that commercial frontier-model traces are distillable but does not evaluate the adaptive attack or the defenses in those frontier-model settings. Thus, the paper convincingly establishes the phenomenon in the studied reasoning setup, but the stronger message that antidistillation defenses generally should be ranked under this adaptive evaluation is not yet supported across substantially different tasks, model families, or release interfaces. Narrowing the scope of that claim would make the evidence and conclusion better aligned.
  • W4: The key step leading from the general defense template to PoE relies on a heuristic value proxy whose relationship to actual distillation value is not directly validated. Sec. 3.2 defines v_gap(x,y) = log pi_ref(y|x) - log pi_stu(y|x) and argues that large teacher-over-student likelihood gaps are plausible carriers of reusable capability. This is intuitive, and the downstream empirical results for PoE are useful, but I could not find an experiment showing that v_gap correlates with the gradient-based value v_grad, with realized per-example learning benefit, or with the examples selected by the adaptive attacker. Consequently, the framework motivates PoE structurally, but it does not yet establish that this particular proxy is a faithful approximation to the learning-value quantity central to the game.
Questions for authors
  • Q1: How should Algorithm 1 be understood as an implementation of the best response in Eq. 6 when Eq. 6 preserves the prompt marginal D and normalizes the exponential tilt separately for each x, whereas Algorithm 1 normalizes weights across different examples in a minibatch? If the intended empirical threat model is actually a KL-constrained change of measure over the joint distribution of (x, y), could you state that formulation and clarify which theoretical result applies to it?
  • Q2: What exact downstream loss L_{theta} is used to compute the gradient-alignment scores in the reported adaptive experiments? In particular, are ground-truth labels from the holdout split used, how many holdout examples contribute to each gradient, how frequently is the gradient recomputed, and is this labeled side information available to the passive baseline in any form? This directly affects how I interpret the attack's information budget and the source of the passive-adaptive improvement.
  • Q3: What is the intended scope of the paper's main evaluation recommendation? Is the claim that adaptive reweighting should generally be expected to weaken antidistillation defenses across tasks and model families, or is the current evidence intended to establish this specifically for reasoning-trace distillation in settings similar to GSM8K/MATH? Q4. Do you have evidence that the likelihood-gap proxy v_gap in Eq. 10 tracks actual example-level distillation value—for example, correlation or rank agreement with v_grad, realized loss reduction after upweighting an example, or the scores used by the adaptive student? Such evidence would substantially strengthen the conceptual link between the minimax framework and the PoE defense rather than treating PoE mainly as an empirically successful heuristic. Q5. For the trace-quality claim, can you clarify how sensitive Fig. 4 is to the choice of judge/rubric and whether the 30-trace human check was sampled across both datasets and all defense strengths represented in the figure? I currently view the evidence as supporting preservation of auditability specifically; a broader validation or a narrower claim would resolve this concern.
2 comments
AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for their careful comments, and we respond to their comments below. We also refer the reviewer to the revised version of the submission we uploaded.

  1. Conditional game versus example reweighting

We agree that a joint student formulation better describes our attack. We revised the submission to let the student choose a joint distribution over prompt-output pairs, and updated the student part of Theorem 3.1 and Appendix B accordingly.

Specifically, writing the released teacher's joint distirbution over prompt-output pairs Prel(x,y):=D(x)πrel(y∣x)P_{\mathsf{rel}}(x,y):=\mathscr D(x)\pi_{\mathsf{rel}}(y\mid x), the student chooses QQ subject to DKL(Q∥∥Prel)≤ρD_{\mathrm{KL}}(Q\|\|P_{\mathsf{rel}})\le\rho. The updated theorem gives $ Q^\star(x,y)= \frac{P_{\mathsf{rel}}(x,y)e^{\eta v(x,y)}} {\sum_{x',y'}P_{\mathsf{rel}}(x',y')e^{\eta v(x',y')}}, $ whenever a finite positive η\eta matches the KL budget. Normalization is now global across pairs, allowing the prompt reweighting used by our attack. The teacher still controls only its conditional output distribution, so its one-sided rule and the PoE derivation remain unchanged.

  1. Downstream loss, auxiliary information, and the source of the gain

We agree that the attacker’s information and scoring procedure should be clear. Given that the reviewer raised several points within the same question, we break the answer down below for clarity:

  • L(θ)\mathscr{L}(\theta) is the student’s completion negative log-likelihood on undefended-teacher traces from disjoint held-out prompts.
  • Labels: No ground-truth solutions or test examples are used to compute the scores.
  • Holdout size: The configured splits contain 2,246 GSM8K and 2,500 MATH prompts; below we add a new experiment where we study the impact of the heldout size of the adaptive attacker for GSM8k dataset; as can be seen, the attacker is still effective with 10 times less holdout samples. the new ablation below uses 2,235 available GSM8K traces.
  • Gradient and scores: Both are computed once, at student initialization, with the gradient accumulated over all available holdout traces.
  • Passive baseline: It trains on the same released traces but does not use the auxiliary holdout set. The holdout data influence the adaptive student through weight selection, rather than direct inclusion in SFT.
  • Additional access and computation: Scoring of examples is a one-time computation on the attacker’s own student and requires no access to teacher or defense-proxy parameters.

2.1. Additional experiments on holdout size

We ran a new ablation suggesting that substantially fewer holdout traces suffice in this setting. We varied the holdout size on GSM8K with ADS λ=0.05\lambda=0.05, β=0.5\beta=0.5, keeping the student initialization, training budget, and 1,209 test questions fixed.

Holdout tracesStudent accuracy
Full: 2,23552.36%
1,00052.52%
50051.20%
25054.18%
10053.10%

With 100 holdout traces, which is less than 2% the total data available to the student for SFT, we observe no significant accuracy drop. Note this is for a single seed so small differences are insignificant.

We can also explain this observation theoretically. Indeed, conditional on fixed released traces and student initialization, if holdout traces are independent and all gradient norms are at most GG, Hoeffding’s inequality gives a uniform score error of at most εk=G22log⁡(2m/δ)/k\varepsilon_k=G^2\sqrt{2\log(2m/\delta)/k} with probability 1−δ1-\delta, where m is the number of released traces and k is the number of holdout traces that the attacker has. For exponential weighting, this changes any pairwise weight ratio by at most a factor exp⁡(2ηεk)\exp(2\eta \varepsilon_k), so estimating scores to a fixed accuracy requires only logarithmic growth in mm.

2.2. Additional details regarding the threat model

Regarding the threat model, we assume that the attacker has a small set of clean examples of the task. We think this is a practical assumption. Indeed, the attacker needs to know which capability it wants to acquire, so it can collect a few clean examples of that task. For example, both GSM8K and MATH come with human-written step-by-step solutions, and a few solutions can also be written by hand. Notice that 100 examples is very small compared to the 5,238 queries that the attacker makes to the teacher. This kind of assumption is also common in the security and privacy literature (Shokri et al., 2017; Carlini et al., 2022).

AuthorsOpenReview ↗

Response (2/2)

  1. Scope of the evaluation recommendation

We acknowledge that our empirical claim concerns reasoning-trace distillation, and we believe this is sufficient for our main recommendation, i.e., passive evaluation alone can overstate the protection of a distillation defense. We are happy to make this more precise if the reviewer has suggestions.

Regarding the choice of datasets: recall that in the introduction, we mention that we follow Savani et al. (2025)’s mathematical reasoning setting to measure the influence of attacker adaptation only, keeping the rest fixed. Regarding teacher/student configurations, we additionally show in Appendix C.4 an adaptive–passive gap with a smaller defense proxy. In terms of other defenses, we also examine the optimized rewritten traces of Ma et al. (2026) in Appendix C on GSM8K, where we find that student accuracy alone can miss a substantial loss in trace quality.

4.Does the likelihood-gap proxy track learning value?

We agree that the likelihood-gap proxy warrants additional empirical validation. We report on a new experiment which scores undefended-teacher traces with Qwen2.5-3B and trains Llama-3.2-3B separately on each score quartile using the same SFT setup. From the lowest to highest quartile, accuracy rises from 47.06±0.6247.06±0.62 to 57.38±0.7257.38±0.72 on GSM8K and from 12.37±0.1512.37±0.15 to 16.93±0.8816.93±0.88 on MATH (three seeds, mean ± standard error). As can be seen, the student accuracy increases with likelihood-gap proxy on both datasets. Finally, it is worth mentioning that a similar likelihood-gap has also been used implicitly in recent work on on-policy distillation in Zhao et al. (2026) and Hübotter et al. (2026).

  1. Judge sensitivity and coverage of the human check

We agree that our empirical evidence supports auditability specifically and that the human-check sampling can be clearer. We ran an additional check covering 60 traces: ten randomly sampled questions per dataset, each with Standard, ADS, and PoE traces from the Table 1 operating points. An independent auditor then used the same rubric, yielding quadratic-weighted (\kappa=0.76) and mean absolute error 0.40 against the automated ratings. We have not tested alternative judges or rubrics, and the human check does not cover all defense strengths; we will state these limits and consistently use “auditability.”

Reviewer_TQ5zOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes PoE which defends against adaptive distillation attacks by redistributing the generations from the teacher model. They show that under an adaptive attack, considering the compute cost and the quality of the reasoning, PoE outperforms ADS. This is because their proposed method redistributes based on forward-pass logits rather than backward-pass gradient calculation.

Critical strengths
  1. The method uses a compute efficient way for the defender to change the distribution of the distilled data from the teacher without computing gradients.
  2. Shows that the effectiveness of PoE increases for adaptive attacks (student).
  3. The ablations compare not only final answer accuracy but also compute cost and reasoning quality, which I believe is more important to understanding the strength of PoE compared to previous attacks (ADS). Especially,
  4. They provide an intuition as to why reasoning is better preserved with PoE (lines 444-448).
Critical weaknesses
  1. Why did the attack do worse in the absence of a defense (57.24 > 56.74)? The attack design does not consider the presence of a defense so it should increase the performance of the learned student regardless of ADS or PoE. It seems like the attack actually failed in its purpose to begin with?
  2. Could you test on harder datasets than MATH500 or GSM8K. These are considered easy math on recent models. I wonder if the same trends hold. I also wonder if the reasoning ADS does better than PoE in the passive setting is because the dataset is easy.
  3. I like table 4 showing that reasoning quality is higher for PoE which is important. Could you provide the mean and standard deviation of the total score?
Questions for authors

See above.

1 comment

We thank the reviewer for their careful comments. We respond to their comments and questions below. We also refer the reviewer to the revised version of the submission we uploaded.

  1. Why does adaptive distillation slightly underperform passive distillation for the undefended teacher?

We do not interpret this small difference as evidence that the attack failed for the undefended teacher. The GSM8K decrease is not statistically significant across three seeds (57.24±0.25 versus 56.74±0.17; (p=0.18)), while MATH shows a small increase (15.17 to 15.29). Our intuition here is that, in the absence of "poisoning" from the teacher, all traces are good and there is little point in reweighting. Consistently, Table 2 flags only 0.02% of standard traces as degenerate, compared with 11.03% for ADS and 4.07% for PoE.

  1. Would the findings hold on harder datasets?

While we agree that additional benchmarks would evidently help assess generality, we chose GSM8K and MATH to measure the influence of attacker adaptation only, keeping the rest fixed in comparison to our main related work (Savani et al., 2025). These tasks also remain challenging for our student, since its untuned accuracy is 0.76% on GSM8K and 1.01% on MATH. Thus, their difficulty for recent models does not say anything about the starting point of the student being distilled here.

  1. I like table 4 showing that reasoning quality is higher for PoE which is important. Could you provide the mean and standard deviation of the total score?

As suggested, we have added the mean and standard deviation of the auditability scores alongside Figure 4:

TeacherGSM8K mean ± SE (SD)MATH mean ± SE (SD)
Standard4.77 ± 0.614.28 ± 0.81
PoE4.55 ± 1.033.88 ± 1.36
ADS4.20 ± 1.103.84 ± 1.13

PoE has a higher mean on GSM8K and a comparable mean on MATH. We will make this distinction clear and describe the measured property as auditability.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal summary

We thank the reviewers for engaging with our paper and for their encouraging comments. In the rebuttal and the revised submission, we responded to their main theoretical and experimental concerns, with the paper’s central contribution remaining the same: passive evaluation can substantially underestimate capability leakage in distillation. The reviewers' comments helped us strengthen this message through additional experiments. In particular, the new results show that the attack remains effective even when the attacker has significantly limited auxiliary knowledge compared to the initial setup in our paper. We also show that the likelihood-gap notion of learning value is predictive of student learning.

In more details, on the theory side, we revised the student formulation to match joint example reweighting, clarified the assumptions of Theorem 3.1, dropped unnecessary parameter-uniqueness claims, and distinguished the one-sided PoE derivation from the full Stackelberg game. We also made the attacker's auxiliary information and scoring procedure clear.

On the empirical side, we added new experiments supporting the likelihood-gap proxy on both datasets, and an ablation suggesting limited sensitivity to holdout traces size. Finally, we reported trace auditability summaries and clarified the human evaluation, and acknowledged the empirical scope and dual-use implications. The new experiments confirm the importance of adaptive evaluation and PoE's competitive cost-leakage trade-off in the studied reasoning-trace setting.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.