acceptodds
Under review as a conference paper at ICLR 2027

CQFA: Channel-Wise Query Feedback for Gated DeltaNet-2

Abstract

Linear attention offers an efficient alternative to softmax attention by enabling sequence modeling with linear-time computation and constant-size recurrent states. Within this framework, compressing sequence history into fixed-capacity memory makes selective state correction essential for retaining useful context while continuously incorporating new information. GDN-2 supports flexible memory updates by decoupling erasure and writing, but its correction remains key-driven and does not explicitly incorporate the query used for memory readout. Q-Delta incorporates query feedback into state correction, yet its per-head scalar coefficient applies the same feedback strength across channels, limiting channel-specific control over state correction. We therefore propose CQFA, which augments GDN-2 with input-dependent channel gates to selectively modulate query-based memory predictions on the erase side, independently of the writing of new values. This design allows feedback to be strengthened in some channels and suppressed in others while preserving the state size and rank-one update structure. We establish a sufficient condition for one-step residual contraction and implement efficient chunkwise parallel training with throughput comparable to that of GDN-2. Systematic analyses and experiments demonstrate that CQFA trains stably and achieves substantial gains over strong baselines on language modeling and information retrieval tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Linear Attention, Gated Delta Networks, Channel-Wise Query Feedback, Associative Memory, Language Modeling, In-Context Retrieval

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_sZfhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CQFA, an extension of Gated DeltaNet-2 that adds query feedback to memory updates through input-dependent channel-wise gates. Unlike scalar feedback, these gates allow different channels to contribute with different strengths, changing both the direction and magnitude of the correction. The method keeps erasure and writing separate and preserves the recurrent memory size. The authors also derive a condition for one-step residual contraction and provide a chunkwise implementation for parallel training.

Experiments compare 340M and 1.7B models with several recurrent baselines on language modeling and zero-shot reasoning. Natural and controlled retrieval are evaluated at 340M. The paper reports higher average task scores while retaining roughly 96% of GDN-2’s training throughput. Additional analyses examine query and key readouts, feedback directions, and local residual behavior.

Critical strengths
  1. The method adds flexibility while preserving the efficient recurrent structure. Channel-wise gates can change both the strength and direction of query feedback without increasing the memory-state size or changing the rank-one update. This is the strongest positive aspect of the proposal because it offers a practical way to make memory correction more expressive.

  2. The core computation is explicit and checkable. The paper provides sequential and chunkwise equations, a reference implementation, and a precise condition for local residual contraction. These details support the basic formulation and provide a useful foundation for implementing and testing the method.

Critical weaknesses
  1. The retrieval results are inconsistent with the stated scoring protocol. Appendix A specifies 384 examples per evaluation set and one binary target per S-NIAH example. The highest nonperfect score would therefore be 383/384, or 99.74%. Yet Table 5 reports scores such as 99.98% for GDN-2 and 99.86% for Gated DeltaNet. Taking the best of three evaluations cannot explain these values. This directly affects the reliability of the retrieval evidence. The actual sample counts, aggregation procedure, and per-example outputs need to be provided before these results can support acceptance.

  2. The main ablation does not isolate the contribution of channel-wise feedback. Appendix B uses a feedback bound of 0.4 for CQFA and 1.0 for Scalar-Q. The comparison therefore changes feedback strength as well as gate granularity. Matching the parameter counts does not remove this confound, and the results cannot establish whether channel-wise control or the different feedback bound explains the gains. Since channel-wise control is the paper’s main contribution, a comparison with identical bounds and comparable tuning is necessary to substantiate its central claim.

Questions for authors
  1. How were Table 5 scores such as 99.98% obtained from 384 single-target examples with binary scoring? Please clarify the denominator and aggregation.

  2. Why were feedback bounds set to 0.4 for CQFA and 1.0 for Scalar-Q? Were both tuned under the same protocol?

  3. Do you have CQFA and Scalar-Q results with identical feedback bounds to isolate the effect of channel-wise control?

  4. How were checkpoints and best results selected? Were all models evaluated on identical examples?

  5. Can you provide results for all three training seeds, including means and uncertainty, to show whether the gains are consistent?

2 comments

Thank you for identifying the reporting inconsistency and the controls needed for the central comparison. We address each question below.

Q1: Retrieval denominator and aggregation. The descriptions of 384 examples and best-of-three selection were retained from the preliminary evaluation. The final results use 28,000 scored examples for each variant/length condition in Table 5 and each task in Table 6. All models use the same examples, with no exclusions. Each condition yields one pooled score under Equation (18). For S-NIAH, this is 100H/28,000100H/28{,}000, where HH counts successful examples. MV/MQ use mean per-example target recall; Table 6 averages its three task scores. The S-NIAH resolution is approximately 0.003571 percentage points, and scores are rounded to two decimals. Values such as 99.98% and 99.86% are therefore compatible with the final sample count. No best-of-three selection is applied. We have corrected the protocol statements and captions; the table values already represent this final pooled evaluation. Exact hit counts and per-example outputs are not supplied in this response.

Q2: Choice and tuning of feedback bounds. The bound Ls=1.0L_s=1.0 follows the (0,1)(0,1) range of Q-Delta's adaptive per-head scalar feedback. Scalar-Q adapts its parameterization to GDN-2 using the same projection parameter count as CQFA. CQFA's L=0.4L=0.4 constrains each coefficient while allowing separate channel contributions. For projected logits near zero, the fixed negative offset favors smaller coefficients. These are design choices made without a shared tuning sweep. See Q-Delta, Appendix B.

Q3: Comparisons with identical bounds. A matched-bound experiment is not included in the current results. We have clarified that the reported Scalar-Q/CQFA comparison is parameter-matched and evaluates the two specified feedback configurations. The common-bound analysis characterizes additional directional flexibility; measuring its independent performance contribution requires common bounds and comparable feedback strengths.

Q4: Checkpoint selection and shared evaluation examples. For the main GDN-2, Scalar-Q, and CQFA comparison, we used three common training seeds under otherwise identical experimental settings. We selected the same seed for which all three training-loss curves were stable, based on inspection of those curves. The 340M task tables use the completed 15B-token checkpoint from each selected run, with one checkpoint per model shared across tasks. We have replaced “best results” with this selection procedure. The final controlled-retrieval evaluation uses identical examples for all models and scores the complete pooled set, separately from training-seed selection.

Q5: Variability across training seeds. Complete per-seed task summaries and uncertainty estimates are not included in the present results. The reported differences therefore describe the selected runs; their consistency across training seeds remains to be evaluated. The similar timing of training instability is an observation about the loss curves, rather than evidence for seed-averaged task performance.

Thank you for the clear reply. Using 28,000 examples makes the reported percentages possible and explains the mismatch with the earlier description. Please include the exact S-NIAH hit counts so readers can check the scores directly.

My main concern remains the comparison between CQFA and Scalar-Q. The feedback bounds differ, and there was no shared tuning process. This means the current results do not separate the effect of channel-wise gates from the effect of feedback strength.

Using the same stable seed for all three models makes the selection clearer. However, results from one selected seed do not show whether the gains hold across runs. Results with the same feedback bounds and scores from all three seeds are still needed to support the paper's central claim.

Reviewer_3MYyOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CQFA introduces channel-wise query feedback into the erase-side memory prediction of GDN-2, leaving the write path independent and the recurrent-state dimensions unchanged; the feedback gates are input-dependent and come from a low-rank projection. On the analytical side, the paper derives a local residual-contraction identity, a chunkwise parallel formulation, and diagnostics for readout geometry. Experiments at 340M and 1.7B show gains in language modeling and in aggregate downstream accuracy, with training throughput at roughly 96% of GDN-2’s. Empirically, the pivotal question is whether channel-wise feedback does better than scalar feedback once the two are compared under conditions that are truly equivalent.

Critical strengths

S1. Technically, the extension is uncomplicated, and it keeps an implementation structure worth keeping: the effective erase direction of Equation (9) still yields a diagonal-plus-rank-one state transition, and Appendix D supplies extensive forward and backward derivations along with reference code.

S2. The scalar-feedback baseline is given the same number of projection parameters as CQFA, a sensible control for the parameters the method adds. GDN-2 and Q-Delta both appear in the experiments, which places the contribution as a particular combination and refinement rather than as the first use of query feedback.

S3. Gains are reported at both scales—Table 3 lists average accuracies of 47.88% and 58.39%—and throughput is measured directly. The local analysis is backed by measured gate and readout statistics with cluster-level uncertainty, so it does not rest on an algebraic possibility alone.

Critical weaknesses

W1. The controlled-retrieval figures do not match the evaluation protocol as described. According to Appendix A.1 there are 384 examples per input seed, and Table 5 reports the best of three seeds; Equation (18) and Appendix A.4 assign a single binary target to each S-NIAH example. Any such percentage therefore has to be a multiple of 100/384, so a score short of perfect cannot exceed 99.7396%. Yet the GDN-2 entry for S-NIAH-1 at 1K reads 99.98%, and other entries are likewise incompatible with that denominator. Taking the maximum over three sets of this kind does not lift the constraint. The effective denominator, the metric, and the aggregation need to be brought into agreement before these results can be read with confidence.

W2. Gate granularity is not isolated by the scalar-versus-channel comparison. In Appendix B, Scalar-Q runs with L_s=1.0 whereas CQFA runs with L=0.4, and on top of that the scalar logits are pooled before the sigmoid. Since the feedback range can affect both stability and optimization, matching parameter counts—valuable though it is—leaves open whether the improvement stems from channel-specific control or simply from a better setting of feedback strength. The fixed-state expressivity comparison of Equation (30) assumes a common bound, so it does not resolve this confound in the experiments.

W3. Because best scores are what get reported, how reliable the task gains are remains unclear. Section 4.1 mentions three training seeds yet reports the best of them, and Appendix A.1 states that a single checkpoint is used across the tables without saying how it was chosen. Means, a measure of spread, and an account of how training-seed selection relates to input-seed selection are needed—all the more so because the aggregate improvements are relatively small. The very small reductions in local residual energy in Appendix E back a diagnostic observation but do not, on their own, establish the downstream gain.

Questions for authors

Q1. Could you share the effective sample counts behind Table 5, together with the aggregation code or raw counts, and explain how a value of 99.98% can arise under the stated protocol of one target per example and 384 examples? If the reported results average over further units, those units should be named and the protocol description revised to match.

Q2. Does the advantage of channel-wise feedback survive when Scalar-Q and CQFA share the same feedback bound and start from comparable gate strengths? A tightly scoped comparison at a common bound goes to the heart of the mechanism being claimed.

Q3. Task summaries per training seed, or a mean with spread, would be welcome, along with the criterion used to pick the displayed checkpoint and a clear separation between training-seed selection and the best-of-three rule applied to synthetic input sets. Under an aggregate with no selection applied, which comparisons still come out favorably?

As things stand, I lean toward rejecting. The method is plausible, the derivations are useful, and the language-modeling improvements reported are encouraging. The main scalar/channel comparison, however, is confounded, and the controlled-retrieval protocol cannot produce numbers of the form that appear in the table—two issues that bear directly on the central claims. A convincing account of the aggregation, together with a comparison at a common bound, could move my assessment. The one-step contraction identity holds when the current gates and targets are held fixed; I do not read it as a global guarantee of stability. Finally, raising the score/protocol discrepancy is a request for technical clarification and in no way an accusation of misconduct.

1 comment

Thank you for examining the evaluation protocol and experimental controls in detail. We have corrected the protocol description and clarified the comparison and run-selection procedure.

Q1 / W1: Effective sample counts and aggregation. The 384-example and best-of-three descriptions belonged to the preliminary evaluation. They were retained when the final results replaced the preliminary scores, causing the inconsistency you identified.

The final evaluation uses 28,000 scored examples for each variant/length condition in Table 5 and each task in Table 6. All models use the same examples, with no exclusions. Each condition yields one pooled score under Equation (18). For S-NIAH, Recall=100H/28,000\mathrm{Recall}=100H/28{,}000, where HH is the total hit count. MV/MQ use mean per-example target recall, and the Table 6 average is the arithmetic mean of its three task scores. Scores are rounded to two decimal places. The S-NIAH resolution is approximately 0.003571 percentage points, making 99.98% compatible with this denominator. No best-of-three selection is applied. The affected protocol statements and captions have been corrected; the table values already represent this final pooled evaluation. Exact hit counts and per-example outputs are not supplied here.

Q2 / W2: Feedback range and gate granularity. The bound Ls=1.0L_s=1.0 follows the (0,1)(0,1) range of Q-Delta's adaptive per-head scalar feedback, with a distinct parameterization adapted to GDN-2 and CQFA's projection parameter count. CQFA uses L=0.4L=0.4 to constrain coefficient magnitude while retaining separate channel control. These choices were made without a shared tuning sweep. See Q-Delta, Appendix B.

The revised text identifies this as a comparison of the specified feedback configurations. Pooling logits before the sigmoid can also affect gate distributions, so parameter matching and the fixed-state common-bound analysis do not isolate gate granularity empirically. A comparison with common bounds and comparable gate strengths remains to be evaluated.

Q3 / W3: Training seeds and checkpoint selection. For the main GDN-2, Scalar-Q, and CQFA comparison, we used three common training seeds under otherwise identical experimental settings. We observed instability with similar timing and selected the same seed for which all three training-loss curves were stable, assessed by inspecting those curves. The 340M task tables use the completed 15B-token checkpoint from each selected run, with one checkpoint per model shared across tasks. Architectural similarity is a possible explanation for the correlated instability, rather than an established cause.

We have replaced the ambiguous “best results” wording with this selection procedure. The tables describe one shared selected seed. Complete per-seed task summaries and an unselected-seed average are not included in the present results, so consistency of the gains across training seeds remains to be evaluated. Evaluation aggregation is separate: each controlled-retrieval score pools all 28,000 examples in its condition.

The revised residual analysis treats the small energy changes as local diagnostics. Its fixed-gate, fixed-target identity describes one-step behavior; the downstream task results provide the evidence for performance of the trained configurations.

Reviewer_Bf4TOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • This paper studies adding query channel feedback to the GDN-2 recurrence, leading to the CQFA architecture

  • The basic idea is to combine the query-feedback from Q-Delta with the per-channel gating of GDN-2. This adds per-channel gating into the query feedback channel

  • They do some experiments showing that query-feedback adds new information to the feedback stream from GDN

  • The resulting algorithm is slightly slower than GDN-2, and is able to reuse the vast majority of its kernel

  • Empirically, CQFA has good performance up to 1.7B params/40B tokens scale

Critical strengths
  • Clear presentation. The paper is well written and easy to follow, with a clear description of the proposed recurrence and its relationship to prior methods.

  • Simple, focused contribution. CQFA extends Q-Delta’s query feedback with channel-wise gates on the GDN-2 backbone. The resulting modification is clean and straightforward, preserving the recurrent state size and rank-one update structure.

  • Useful analysis of query feedback. I appreciate the effort to examine readout geometry, feedback direction, and local residual dynamics. These analyses provide a useful starting point for understanding the method, although their connection to the downstream gains could be strengthened, as discussed in the weaknesses.

  • Strong empirical results with modest overhead. CQFA achieves strong overall results across language modeling, reasoning, and retrieval tasks. The efficient kernel implementation retains most of GDN-2’s training throughput, making the additional computational cost modest.

Critical weaknesses
  • Limited explanation of the benefit of channel-wise feedback over scalar feedback. Much of Section 3.1 motivates incorporating query feedback by showing that query and key readouts can differ at individual tokens, despite occupying largely similar dominant subspaces. This primarily supports the motivation behind Q-Delta, rather than explaining CQFA’s additional benefit from channel-wise gating. Section 3.2 demonstrates that channel-wise gates change the feedback direction, but why these changes improve performance remains unclear. Beyond the task-level ablations, a matched comparison of Scalar-Q and CQFA would strengthen the analysis: how do their learned readout geometry, feedback–residual alignment, and resulting memory corrections differ? This would help connect the additional directional flexibility to its practical benefit.

  • The local residual analysis does not establish the mechanism behind the downstream gains. Section 3.3 reports only small reductions in aggregate pre-update residual energy (0.036–0.100%). These results do not rule out residual improvement as a contributing mechanism, but they leave unclear whether it explains the performance gains. Moreover, contraction of the current mixed prediction residual does not establish that the update preserves useful associations learned at earlier timesteps. A finer-grained, sequence-level analysis would be valuable: as new key–value pairs arrive, how does the error on previously fitted associations evolve? Does CQFA reduce interference or improve retention relative to GDN-2 and, especially, the parameter-matched Scalar-Q control? Measuring the tradeoff between correcting the current association and disturbing previous associations would provide a stronger explanation of CQFA’s success than demonstrating local contraction and increased update flexibility alone.

Questions for authors
  • What motivated the choices L=0.4L=0.4 and δλ=−0.9\delta_\lambda=-0.9? How sensitive are performance and stability to the feedback bound and logit offset?

-Could you expand the efficiency evaluation to include throughput across head dimensions and autoregressive decoding speed, compared with GDN-2 and other baselines?

2 comments

Thank you for the suggestions on connecting directional flexibility to memory correction. We have expanded this explanation and clarified what the analyses measure.

  1. From channel-wise feedback to memory correction. Section 3.1 motivates query-dependent feedback, which is also the motivation for scalar query-feedback methods. CQFA adds control over the relative contributions of query channels: a scalar scales their combined contribution, whereas channel-wise reweighting can also change its direction. The measured orthogonal feedback component shows that the trained model uses this additional freedom.

    Equation (9) makes its effect on a fixed probe xx explicit: (St−Sˉt)⊤x=(kt⊤x)rtq(S_t-\bar S_t)^\top x=(k_t^\top x)r_t^q, relative to the decayed pre-update state Sˉt\bar S_t. Key overlap determines the coupling, while the gates modify rtq=rt0−Mtλtr_t^q=r_t^0-M_t\boldsymbol{\lambda}_t, changing the magnitude and value-space direction of the correction. Channel reweighting can thus influence both current fitting and earlier readouts. Whether an earlier association improves depends on the correction's alignment with its existing error. This relation, added in Appendix E.3, motivates a mechanism for studying correction and interference; improved retention requires sequence-level evidence.

  2. Interpretation of the local analysis. The observed 0.036-0.100% residual-energy reductions measure a local effect. Section 3.3 now presents them as diagnostics, while the downstream experiments assess the trained configurations' task performance. The contraction identity concerns the current fixed target and gates. Matched Scalar-Q/CQFA geometry and alignment measurements, together with a sequence-level comparison of earlier-association error, would connect directional flexibility to correction and retention more directly. These comparisons are not included in the current results.

  3. Feedback hyperparameters. For projected logits near zero, the offset δλ=−0.9\delta_\lambda=-0.9 favors smaller coefficients. At zero logits, 0.4σ(−0.9)≈0.1160.4\sigma(-0.9)\approx0.116. This motivates a modest initial coefficient scale, while the projection learns how feedback adapts during training. The bound L=0.4L=0.4 caps each coefficient while preserving channel-specific control. These are design choices; sensitivity to the bound and offset has not been measured in a sweep.

  4. Efficiency and parallel implementation. Our efficiency contribution is to introduce channel-wise query feedback while preserving GDN-2's efficient execution structure. CQFA retains the recurrent-state dimensions and diagonal-plus-rank-one transition. Section 3.4 and Appendix D derive the chunkwise forward and backward computations, showing how the added feedback fits into the existing factorization without enlarging the triangular system or recurrent state. The measurements complement this analysis: across sequence lengths from 2K to 16K, CQFA retains 95.99-96.34% of GDN-2's training throughput. Together, the derivation and measurements support the practical efficiency of this focused architectural extension. The current evaluation characterizes parallel training; varying head dimensions and profiling autoregressive decoding are natural extensions of the same implementation.

Thanks for the response. I will be maintaining my score (3) after reading this and the other reviews.

Reviewer_F8pWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Channel-wise Query Feedback Attention (CQFA), extending Gated DeltaNet-2 with input-dependent, per-channel query feedback for memory correction. Unlike scalar feedback, CQFA adjusts both the magnitude and direction of query contributions while preserving independent erase/write gating, recurrent-state size, and rank-one updates.

The authors derive a conditional one-step residual contraction criterion, analyze key/query readout geometry and feedback behavior, and develop an efficient chunkwise implementation. Experiments show improvements in language modeling and reasoning at both 340M and 1.7B scales, alongside retrieval gains at 340M, while maintaining approximately 96% of GDN-2's training throughput.

Critical strengths
  • Efficient structural extension. CQFA preserves GDN-2's recurrent-state size and diagonal-plus-rank-one updates, achieving approximately 96% of its training throughput. The complete chunkwise derivation supports practical implementation.

  • Promising empirical results across scales and tasks. CQFA achieves the highest eight-task average at both model scales and leads aggregate natural-context and multi-association retrieval results (Tables 3–6). Non-scalar feedback analysis further supports the proposed mechanism.

Critical weaknesses
  • Unmatched feedback bounds confound the central comparison. CQFA uses L = 0.4, while Scalar-Q uses L_s = 1.0 (Appendix B). Since excessive feedback can increase residual error (Eqs. 12, 39–40), the comparison does not isolate the benefit of channel-wise gating. Matched-bound experiments and realized feedback-strength distributions are needed.

  • RULER scores are inconsistent with the stated evaluation protocol. Appendix A.4 defines single-target S-NIAH accuracy over 384 examples, implying increments of approximately 0.2604 percentage points (100/384). However, Table 5 reports 99.98% for GDN-2 (1K S1), which is impossible under this protocol, even with best-of-three selection. The authors should clarify the actual denominators, hit counts, and aggregation procedure.

Questions for authors
  • Does channel-wise gating still improve performance under matched feedback settings? Why does Scalar-Q use Ls=1.0L_s=1.0 while CQFA uses L=0.4L=0.4, and how were these bounds selected? Please compare both methods under matched bounds (0.4 and 1.0), keeping other settings consistent, and report the realized feedback-strength distributions. These results are essential for isolating the contribution of channel-wise gating.

  • How are the RULER scores computed and aggregated? Table 5 reports 99.98% for GDN-2 (1K S1), which is inconsistent with accuracy computed over 384 single-target examples per evaluation set. Please clarify the actual sample counts, hit counts, and aggregation procedure, including best-of-three selection. Do the reported retrieval gains remain under a consistent evaluation protocol?

1 comment

Thank you for identifying the controls and protocol details needed to interpret the results. We have clarified both points in the revision.

  1. Feedback bounds and the Scalar-Q comparison. The choice Ls=1.0L_s=1.0 follows the (0,1)(0,1) range of Q-Delta's adaptive per-head scalar feedback. Scalar-Q adapts this feedback to GDN-2 using a low-rank projection matched to CQFA's parameter count. CQFA uses L=0.4L=0.4 to constrain each coefficient while allowing separate control of channel contributions. For projected logits near zero, the fixed offset δλ=−0.9\delta_\lambda=-0.9 favors smaller coefficients; the projection learns their subsequent adaptation. These are design choices, without a shared hyperparameter sweep. Scalar-Q's parameterization is distinct from Q-Delta's, as clarified in Appendix E.3. See Q-Delta, Appendix B.

    We have described the experiments as a parameter-matched comparison of the specified configurations. The common-bound analysis characterizes the additional directional freedom of channel-wise gates. Isolating its empirical contribution requires comparisons at common bounds of 0.4 and 1.0, together with realized feedback-strength distributions; these measurements are not included in the current results.

  2. RULER scoring and the reported percentages. The 384-example and best-of-three descriptions came from the preliminary evaluation and were retained when the final evaluation results replaced the preliminary scores. We have corrected the protocol text and captions.

    The final protocol uses 28,000 scored examples for each variant/length condition in Table 5 and each task in Table 6. All models receive the same examples, with every example included. Each condition produces one pooled score using Equation (18). For S-NIAH, each example has one target, giving 100H/28,000100H/28{,}000, where HH is the total hit count. MV/MQ average per-example target recall; the Table 6 average is the arithmetic mean of its three task scores. Scores are rounded to two decimal places. The S-NIAH resolution is approximately 0.003571 percentage points, making 99.98% compatible with the final denominator. No best-of-three selection is applied.

    Tables 5 and 6 report this final pooled evaluation, so the correction changes the protocol description rather than the table values. Exact hit counts and per-example outputs are not supplied in this response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.