acceptodds
Under review as a conference paper at ICLR 2027

Token-Adaptive Concrete Score Distillation (TA-CSD)

Abstract

Knowledge distillation for autoregressive language models commonly applies a fixed temperature transform to logits before computing a divergence. We begin with an analytical observation: for any objective that depends on logits solely through pairwise differences, including softmax divergences over a fixed candidate window and concrete-score objectives, the centering component of logit standardization is mathematically inert. Consequently, standardization operates entirely as a per-token temperature , collapsing the question of whether to standardize into which estimator of that per-token scale to use. The conventional plug-in choice, the student's head standard deviation , is blind to teacher confidence, and because concrete-score gradients scale inversely with temperature, plug-in standardization silently acts as an effective learning rate modifier. We empirically quantify this effect in language model distillation, offering a candidate explanation for conflicting reports on logit standardization in the literature. Token-Adaptive Concrete Score Distillation (TA-CSD) predicts the per-token temperature via a lightweight multi-layer perceptron (MLP) gate conditioned on detached supervisory and predictive uncertainty features. In practice, the gate is trained jointly with the student under score distillation with detached features, while bilevel soft-ECE optimization provides a theoretical and diagnostic foundation. Under its learned scale, TA-CSD remains effective-learning-rate neutral (measured ratio , with mean ), so its differences from unscaled distillation cannot be attributed to step size. Evaluating across multiple model families and seeds with paired example-level bootstrap tests reveals a dissociation between calibration error and task capability. Plug-in standardization lowers top-1 calibration error only nominally ( vs. for CSD, not significant) while hedging (flattening the predictive distribution), losing of decisive () token predictions, degrading mathematical reasoning on GSM8K by percentage points (), and significantly worsening proper score (). While on-policy rollouts (GKD) achieve strong calibration, they alter the training distribution and demand over more training compute. Among offline cached distillation methods, TA-CSD avoids this harm: our clamped formulation attains the best point-estimate Brier score () and protects GSM8K reasoning (, vs. for SWS), while a residual gate reaches smECE at the cost of reasoning ().

Primary area:
foundation or frontier models, including LLMs
Keywords:
knowledge distillation, speculative decoding, large language models, temperature scaling, model calibration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QfcAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the role of logit standardization in concrete score distillation for autoregressive language models. It argues that, for a fixed candidate window and a shift-invariant score objective, subtracting the logit mean is inert and the operational effect of standardization is a token-wise temperature. Based on this observation, the paper proposes Token-Adaptive Concrete Score Distillation, or TA-CSD, which predicts a student-side temperature from detached teacher and student uncertainty features. Experiments across multiple same-tokenizer teacher-student pairs evaluate calibration, proper scoring rules, token accuracy, and downstream reasoning. Additional cross-tokenizer experiments examine the effect of fixed standardization regimes across several divergence formulations.

Critical strengths
  1. The paper provides a clear analysis of logit standardization in concrete score distillation. In particular, the authors distinguish the roles of centering and scaling in the shift-invariant score objective and explain that what actually affects the optimization process is token-dependent scaling. This analysis helps to explain why different standardization strategies may exhibit different effects under different objectives and training settings.
  2. The experiments do not rely solely on a single calibration metric but report calibration, sharpness, token accuracy, and reasoning performance simultaneously. The paper also includes multi-seed results, paired comparisons, matched-learning-rate control, and experiments on the second model family, allowing readers to observe more comprehensively the benefits and costs brought by the method.
Critical weaknesses
  1. The paper proposes a bilevel calibration objective, but the TA-CSD used in the main experiment employs a jointly trained gate based on detached features. The relationship between the optimization goals and the actual training process of the two has not been fully explored, which may make it unclear to readers whether the bilevel formulation is the main part of the method, a theoretical explanation, or a future extension direction.
  2. The paper uses effective-learning-rate to explain the impact of scaling on the training dynamics, but this metric is mainly based on the proportion of logit gradients. For training processes using adaptive optimizers such as AdamW, the logit-gradient ratio is not necessarily equivalent to the actual parameter update magnitude. Therefore, related conclusions should be further explained in conjunction with parameter-update statistics or normalized update norms.
  3. Current experiments have shown that the method's effectiveness varies with model settings and tokenizer configurations, but the direct verification of TA-CSD in cross-tokenizer or more significant distribution variation scenarios is still limited. Supplementing these settings can help illustrate whether the learned token-specific scale has stable transferability, rather than relying mainly on specific model combinations.
Questions for authors
  1. Can the author more clearly distinguish the positioning of the jointly trained gate and the bilevel calibration objective? For example, what is the actual optimization target in the main experiment, and is the bilevel formulation used for diagnosis, theoretical explanation, or future extension?
  2. Could the authors evaluate TA-CSD itself, rather than only fixed standardization variants, on at least one cross-tokenizer setting? Reporting calibration, Brier score, token accuracy, and downstream reasoning together would clarify whether the learned gate transfers beyond same-tokenizer distillation.
2 comments
AuthorsOpenReview ↗

Response to Reviewer QfcA

We thank the reviewer for their careful reading and the positive assessment. All changes are colored in blue in the revised PDF, so below we point to them rather than repeat them. "New" marks results produced for this response.

W1 / Q1. Is the bilevel objective the method, a theoretical explanation, or a future extension? (new) The submission was unclear here. The revision fixes this in the abstract, §3.4, Tables 3–4 and Appendix G.

  • The method is joint training. The gate is trained jointly with the student on L_score + λ_CE·L_CE, with detached features. Every TA-CSD number in the main results uses this objective, and no calibration metric enters training.
  • Bilevel is a tested alternative, and it lost. We ran full-scale bilevel training: a one-step finite-difference unroll, with a soft-ECE outer loss on the training micro-batch, over 3 seeds. It preserves GSM8K but calibrates worse than joint training (smECE 0.109 vs 0.055). Adding a Brier term to the outer loss helps only slightly.
  • What we take from it. Noisy micro-batch hypergradients are a worse training signal for the gate than the distillation loss itself.

W2. The logit-gradient ratio is not the parameter update under AdamW. (new) We agree, and measuring the update changed our interpretation (§5.1, new Appendix B, Table 6):

  • Step size. We measured each arm at its own trained checkpoint with its own AdamW state. Its step is 1.11× (SWS), 1.01× (AdaKD-head) and 1.03× (TA-CSD-C) that of CSD.
    • Adam absorbs the raw 3.2× attenuation of the plug-in arms, and gradient clipping never fired for them.
    • The raw ratios (Table 2) are now described as an effective-LR change only for SGD-like updates.
  • Direction. What persists is the direction of the update. The plug-in arms shift it toward the CE term: at initialization, the score-to-CE gradient ratio is 3.8 for SWS vs 11.3 for CSD. TA-CSD-C keeps CSD's balance.
  • Exact mechanism for SWS. Because SWS differentiates through σ̂^S, the score term is invariant to the student's logit scale (unit test added). Sharpness is therefore set by the label-smoothed CE term, which caps confidence near 0.9.
    • Retrained without label smoothing (new, seed 0), SWS flips from under- to over-confident, which confirms the mechanism.
    • Its GSM8K recovers only partially (App. B).
  • Matched-LR retraining. This also explains why it fails: under Adam it enlarges an already comparable step, without restoring the missing gradient along the logit scale.

W3 / Q2. TA-CSD on a cross-tokenizer pair. (new) We ran it once, and the result is negative (App. C, Table 8).

  • Result. On Qwen2.5-7B → Llama-3.2-1B with rank-sorted alignment, the gate hurts both calibration and GSM8K (14.7% vs 27.8% for CSD), even though training was stable.
  • A possible cause, which we have not tested. After alignment, the gate's teacher-side features may no longer refer to the student's own tokens.
  • Scope. Limitation (iv) states this result, and all TA-CSD claims remain same-tokenizer claims.
  • Cross-family evidence. The evidence we do have is Pair B (Table 4, §5.4). There, TA-CSD-C roughly halves CSD's calibration error and gives the best GSM8K among distilled arms. SWS calibrates better but has the lowest GSM8K, which the paper now describes as a trade-off, not dominance.
Reviewer_QfcAOpenReview ↗

To author

Thank you for your clarification. Most of my concerns have been addressed. I will maintain the 3 score.

Reviewer_7mevOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper suggests using a token-adjusted concrete score matching for knowledge distillaiton. This paper identifies that the temperature adjustment would only be relevant to the concrete score matching, so the mean shift would be unnecessary. The paper suggests two modes of scale parametrization with a learnable feature.

Critical strengths

Good and intuitive point that likelihood ratio matching does not get benefit from mean shift and matching.

Critical weaknesses

This paper suggests using a token-adjusted concrete score matching for knowledge distillaiton. This paper identifies that the temperature adjustment would only be relevant to the concrete score matching, so the mean shift would be unnecessary. The paper suggests two modes of scale parametrization with a learnable feature.

It is expected that the concrete score matching, which is a matching of likelihood ratio, would take only the scaling effect, not the mean shift effect. This explains Proposition 1 and Corollory 1 in its simple term. This might interests some readers, but the intuition itself is very straightforward.

The remedy for the token-adapation is simply introducing a learnable term in determining \tau, which is the scaling parameter. Two stylized adaptation is suggested, which seems to be designed without specific theoretic background in its function design, eq 8 and 9. The remedy is very simple, as well.

The result section compares the suggested token-adaptation of CSD to the fixed scaling of CSD. The reported CSD is using a fixed \tau of 1, which is arbitrarily chosen, and this is not a fair comparison. CSD should have given a chance to calibrate its \tau. \tau needs to be further analyzed with diverse \tau settings. Additionally, what would be the value of TA-CSD when TA-CSD-C and post-hoc TS are almoste similar in their peformances.

The result seems that this token-adaptation makes good change toward performances. The only difference between CSD and TA-CSD is the token-adaptiveness. Is this only reason for the improvemen? \tau tends to be closer to 1.0, which means that token-adaptation prefers the sharpnes in the token selection. So, \tau may have been overfitted to some word choices, which may improve some accuracy metric by sacrificing some diversity in generation tasks. I am interested in per-token \tau average distributions. Any specific types of tokens with high \tau?

"...in Eq. equation 2 with...." obviously writing error.

Questions for authors

see the weakness section

1 comment
AuthorsOpenReview ↗

Response to Reviewer 7mev

We thank the reviewer for their time and effort. All changes are colored in blue in the revised PDF, so below we point to them rather than repeat them. "New" marks results produced for this response. Pair A is Qwen2.5-7B-Instruct → Qwen2.5-0.5B, with the full GSM8K test set.

W1. The reduction is intuitive. We agree. The inertness of centering follows directly from shift invariance, and we present it as a short observation rather than a deep result. Its value is practical in two ways:

  • It turns "should we standardize?" into "which per-token scale should we use?".
  • It says what must be controlled when standardization is compared with plain distillation: the scale, and through it the gradient gain.

The gate is deliberately simple (177 parameters). The contribution is less the remedy itself than the controlled comparison it enables: the same objective with plug-in, fixed, mean-matched, learned-global and token-adaptive scales, evaluated with proper scores and reasoning.

W2. CSD with τ = 1 is an arbitrary, untuned baseline. (new) This is a fair point, and we now tune it. The results are in Table 3, the new "Adaptivity vs. level" paragraph in §5.2, and Appendix C (Table 7).

  • Training-time sweep (1 seed). CSD with a constant τ ∈ {0.75, 1, 1.14, 1.25, 1.5, 1.86} calibrates best at τ = 1.14–1.25, and GSM8K falls as τ grows. On the same seed, TA-CSD-C beats every constant on smECE, Brier, NLL and GSM8K.
  • Mean-matched constant τ = 1.14 (3 seeds). It improves on CSD. Unscaled, it is statistically indistinguishable from TA-CSD-C (Brier p = 0.67). A tuned constant therefore does recover TA-CSD-C's unscaled calibration, and the paper now says so (Limitation v).
  • Where adaptivity shows. The students are under-confident. The constant raises the overall confidence level, while TA-CSD-C keeps CSD's level and redistributes confidence across tokens. Post-hoc TS fixes the level. After TS, all 5 TA-CSD-C seeds beat all 3 constant seeds (smECE 0.042 vs 0.061, p ≤ 0.036).
  • Post-hoc τ. Fitted on a disjoint split, τ* is below 1 for every model (0.78 for CSD). TS therefore sharpens the models, consistent with under-confidence (App. C).

W3. What is the value of TA-CSD if post-hoc TS performs about the same? (new) The two are complementary, not substitutes. We applied the same TS to TA-CSD-C (Table 3, §5.2):

  • TA-CSD-C + TS reaches smECE 0.0422 and Brier 0.1592, vs 0.0602 and 0.1630 for CSD + TS.
  • All five TA-CSD-C + TS seeds beat all five CSD + TS seeds on smECE, Brier and NLL (exact permutation p = 0.008, the smallest attainable with 5 vs 5).
  • This comes close to on-policy GKD (0.0399 / 0.1587) without student rollouts, and GSM8K is unchanged.

W4. Is token adaptivity the only difference, and does τ overfit to particular words? (new) All arms share the data, cache, score space, CE term, optimizer and schedule, and differ only in the scale. We now analyse the τ the gate emits on 136,565 held-out tokens from all five checkpoints (Fig. 2, §5.2 "What the gate learned"):

  • It tracks uncertainty. τ spans [0.71, 2.86], correlates with teacher entropy (r = 0.84), and is nearly unrelated to position.
  • Which tokens get high τ. The highest scales go to tokens where the teacher is uncertain (1.52 in the top teacher-entropy quartile, vs 0.84 in the bottom one) and to tokens the student gets wrong (1.35 vs 1.00).
  • It is not tied to particular words. By token type, content-word starts get 1.26 and function words 1.18, while digits (0.96) and subword continuations (0.95) get sharper scales.
  • It does not trade accuracy for calibration. Token accuracy and GSM8K match CSD. smECE on two random halves of the held-out set differs by much less than seed-to-seed variation (App. F).
  • Adaptivity adds over a constant at the same mean only after recalibration (W2).

Typo. The duplicated "Eq. equation" strings are fixed.

Reviewer_n6JFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how logit standardization affects concrete-score distillation for autoregressive language models. Its starting observation is that centering cancels in shift-invariant objectives, leaving temperature scaling as the relevant transformation. It proposes TA-CSD, a small gate trained jointly with the student to predict token-specific temperatures from detached uncertainty features.

The evaluation compares clamped and residual gates with unscaled distillation, plug-in standardization, temperature controls, and other distillation approaches across model families and seeds. The reported results highlight trade-offs between calibration and reasoning: clamped gating mitigates some degradation from standardization, while its advantages over unscaled CSD and post-hoc temperature scaling are less conclusive.

Critical strengths

1. The analytical framing is useful.

The observation that centering cancels in shift-invariant distillation objectives helps clarify what standardization actually changes. Section 3.1 also distinguishes treating the scale as detached from differentiating through normalization. This provides a useful foundation for examining temperature choices and their optimization effects.

2. The evaluation acknowledges meaningful trade-offs.

I appreciate the joint reporting of calibration, scoring quality, and downstream reasoning across model families and seeds. Tables 3–4 show why an improvement in calibration error alone can be misleading. The authors also acknowledge that the main-pair gains over CSD are not significant, that post-hoc temperature scaling performs comparably, and that distilled models can remain worse than the base student. These qualifications make the reported findings more informative.

Critical weaknesses

1. The supplied implementation does not clearly match the method and baselines described in the paper.

Section 3.3 specifies a sigmoid gate and global training progress as a feature. However, the supplied configuration selects a hard-clipped log-temperature mapping, and the implemented progress feature is token position within the sequence. The supplied AdaKD baseline also uses student top-eight probability mass, whereas the manuscript describes a teacher–student Hellinger-difficulty rule. These differences affect both the proposed method and a key comparison. The authors need to identify the exact code and configurations behind the reported results and reconcile these discrepancies; the supplied defaults alone do not establish how the historical experiments were run.

2. The evidence does not yet isolate the benefit of token-level adaptation.

The fixed-temperature control uses 1.86, while the learned gate’s reported mean is 1.14. As the paper acknowledges, this comparison does not separate adaptivity from the overall temperature scale. A suitably tuned or jointly learned global-temperature baseline would make the claimed benefit more convincing, particularly given the competitive post-hoc scaling results in Table 3.

I also find the “effective learning rate” interpretation too strong. The supplied diagnostic measures gradients with respect to logits with the auxiliary cross-entropy loss disabled; it does not measure actual AdamW parameter updates under the full training objective. A gradient-norm ratio close to one therefore does not establish that optimization-scale effects have been ruled out. This weakens a central explanation offered for TA-CSD’s behavior.

I would raise my score if the authors reconcile the implementation discrepancies and substantiate the claimed advantage with faithful baselines and appropriate temperature and optimization controls.

Questions for authors
  1. Which exact code revision and configurations produced the main tables? Please clarify the differences in gate parameterization, the training-progress feature, and the AdaKD baseline, and indicate whether these differences are confined to the released implementation.

  2. Does TA-CSD retain a meaningful advantage over a suitably tuned or jointly learned global temperature under the same training budget? This comparison would directly affect my assessment of the contribution from token-level adaptation.

  3. What supports the claim of effective-learning-rate neutrality under the actual optimizer and complete training objective? Can the authors provide parameter-update measurements, or clarify which conclusions remain justified if the existing diagnostic is interpreted only as a logit-gradient comparison?

2 comments
AuthorsOpenReview ↗

Response to Reviewer n6JF

We thank the reviewer for their time and effort. We additionally thank the reviewer for checking the released code against the paper. You were right on all three discrepancies. They were errors in the paper's description, not in the experiments, and no reported number changes. All changes are colored in blue in the revised PDF, so below we point to them rather than repeat them. "New" marks results produced for this response.

W1 / Q1. Exact code and configurations behind the tables.

  • Provenance (not in the paper). Every run wrote a run_manifest.json with its resolved config and git revision:
    • Pair A seeds 0–2: bb0f808.
    • Matched-LR arms and τ = 1.86: 4c4c94b.
    • Seeds 3–4, Pair B, residual gate, E4, and E8 seeds 0–2: 7828e3e.
    • Bilevel, SelecTKD, and E8 seeds 3–4: 34dc9a6.

We diffed the gate, score-space, objective, adapter and trainer code across these commits. The forward path of every reported arm is identical, and later commits only add options.

  • Gate (§3.3, Eqs. 8–10, Table 5). Every TA-CSD-C run uses the hard log-space clip, initialized at τ = √2. The sigmoid form that the submission wrote as Eq. 8 is now labelled as an untrained alternative.
  • Features (§3.3). The "training-progress" feature is actually token position. h/H is 0 for every student, since all are single-head, so the gate effectively uses 8 features. Position barely matters (r = 0.005 with τ).
  • AdaKD (§4). The reported arm is a student head-mass rule, now relabelled AdaKD-head throughout. We implemented the Hellinger rule and report its gradient scale (Table 2, new). We did not retrain it, so we make no calibration or capability claim for it.
  • Supplement. It is regenerated and now includes the configs of every new run, the sweep-table script and the scale-invariance unit test.

W2 / Q2. A tuned or jointly learned global temperature. (new) The τ = 1.86 control came from a logged mean that also averaged prompt and padding positions. Over supervised tokens, the gate's mean is 1.14 (§4). We trained both controls you asked for (Pair A, 3 seeds each, full schedule), plus a CSD temperature sweep. The results are in Table 3, a new "Adaptivity vs. level" paragraph in §5.2, and Appendix C (Table 7). In short:

  • Unscaled, a constant τ = 1.14 matches TA-CSD-C (Brier p = 0.67). You were right that this needed testing, and Limitation (v) now says so.
  • The two get there by different routes. The constant shifts the overall confidence level. TA-CSD-C keeps CSD's level and reallocates confidence across tokens.
  • Adaptivity shows after recalibration. Post-hoc TS corrects the level. After TS, all 5 TA-CSD-C seeds beat all 3 constant seeds on smECE, Brier and NLL (exact permutation p ≤ 0.036).
  • A learned global τ drifts to 0.96 and ends up worse than CSD.
  • Sweep (1 seed). Calibration is best at τ = 1.14–1.25. On the same seed, TA-CSD-C beats every constant on all four metrics.

W3 / Q3. Effective-LR neutrality under the actual optimizer. (new) We agree that the logit-gradient ratio does not settle this. Measuring the actual update changed our interpretation (§5.1, new Appendix B, Table 6):

  • Step size. Plug-in standardization attenuates the raw gradient 3.2× (Table 2), but that holds only for SGD-like updates.
    • Under AdamW, with each arm's own trained state, the SWS step is 1.11× CSD's. This is also why matched-LR retraining did not help.
    • Gradient clipping never fired for the plug-in arms.
    • "Effective-LR neutral" now refers only to raw gradients.
  • Direction: an exact mechanism. SWS differentiates through σ̂^S, so its score term is invariant to the student's logit scale. The gradient along that direction is exactly 0 (unit test added).
    • Sharpness is then set only by the label-smoothed CE term, which caps confidence near 0.9, the band SWS hedges into.
    • AdaKD-head fails the other way and becomes over-confident.
    • TA-CSD-C keeps CSD's balance between the score and CE terms.
  • Direct test (new, seed 0; App. B). Retrained without label smoothing, SWS flips from under- to over-confident, which confirms the mechanism. Its GSM8K recovers only from 22.5% to 25.6% (CSD: 35.9%). The mechanism therefore explains the hedging, but not most of the capability loss.

Thank you for the additional experiments. The tuned-temperature and post-hoc calibration comparisons are useful and strengthen the empirical case. However, my main concerns about implementation consistency, baseline fidelity, and the effective-learning-rate interpretation remain unresolved. I am therefore maintaining my current rating, but remain open to revising it if these issues are addressed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We sincerely thank the reviewers for their time commitment and insights. We appreciate the thorough and concise reviews.

The reviewers agree on what is useful in the paper: the reduction of standardization to a per-token scale, and reporting calibration, proper scores and reasoning together. We have uploaded a revision with every change shown in the color blue. Each reply points to where its points are addressed, and this summary lists the main changes.

1. Code–paper discrepancies (n6JF). These were errors in describing the experiments, not in running them, and no reported number changes.

  • The gate is a hard log-space clip (§3.3, Eqs. 8–10).
  • The "training-progress" feature is token position, and the depth coordinate is constant (§3.3).
  • The "AdaKD" arm is a student head-mass rule, relabelled AdaKD-head (§4).

2. Post-hoc temperature scaling (7mev, n6JF). (new) TS and TA-CSD are complementary. TA-CSD-C + TS beats CSD + TS on every seed (smECE 0.042 vs 0.060, p = 0.008) and comes close to on-policy GKD at offline cost (Table 3, §5.2).

3. "Effective learning rate" under AdamW (n6JF, QfcA). (new) This changed our interpretation (§5.1, Appendix B, Table 6).

  • Under AdamW, the raw 3.2× attenuation of plug-in standardization does not become a smaller step.
  • What persists is the update's direction. SWS's score term is exactly invariant to the student's logit scale, so its sharpness is set by the label-smoothed CE term, the band SWS hedges into.
  • Retraining SWS without label smoothing confirms this mechanism.

4. Bilevel objective (QfcA). (new) The method is joint training. Full-scale bilevel training was run and calibrates worse, and it is now presented as an evaluated alternative (§3.4, Appendix G).

5. Adaptivity vs. level (n6JF, 7mev). (new) We trained a mean-matched constant τ = 1.14, a learned global τ (3 seeds each) and a CSD temperature sweep (Table 3, §5.2, Appendix C).

  • Unscaled, the constant matches TA-CSD-C.
  • After post-hoc TS, TA-CSD-C beats the constant on every seed (p ≤ 0.036).
  • Limitation (v) states this.

6. Cross-tokenizer (QfcA). (new) In one run the gate hurt (Appendix C, Table 8, Limitation iv). All TA-CSD claims remain same-tokenizer claims.

Not done. We did not retrain a Hellinger-difficulty AdaKD, and we make no claim about it beyond its gradient scale (Table 2).

Other corrections.

  • The τ = 1.86 control was set from a statistic that included prompt and padding positions. Over supervised tokens, the gate's mean is 1.14 (§4).
  • E[1/τ] ≈ 0.90, not 1 (§5.1).
  • A τ-distribution and token-type analysis is added (Fig. 2).
  • The supplement includes the configs of every new run.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.