acceptodds
Under review as a conference paper at ICLR 2027

Lead into Gold: The Surprising Effect of Symbolically Generated Deduction Data on LLMs

16 views

Abstract

Synthetic data now supplies much of the instruction tuning of language models, and it is usually written by another language model with the goal of sounding as natural as possible. We ask whether data at the opposite extreme also helps, a corpus of multi-hop deduction problems and their solutions that a short program generates with no model in the loop. The documents are correct by construction, but they carry no world knowledge and they name their constants abstractly. We replace a small fraction of a standard instruction-tuning mixture with such documents while holding the number of training examples fixed. On Qwen2.5-7B, replacing 5 to 10 percent of the mixture raises accuracy by 14 points on ProofWriter, 6 points on FOLIO, 7 points on the multi-step subtasks of BIG-Bench Hard and up to 7 points on the quantitative items of GPQA-Diamond, and it leaves general benchmarks unchanged. These benchmarks cover formal logic, language inference, word problems and the sciences, so the gain is not confined to the domain of the added documents. The gain on ProofWriter holds on models of 3 families and from 3B to 32B parameters. As little as 1 percent of the mixture teaches the model to write long derivations that a symbolic checker accepts, an ability that no baseline model in our study has. Two mechanisms account for the gains. The model learns to carry an intermediate conclusion forward through a chain of steps instead of restarting from the premises, and it stops preferring one answer label over the others, a bias we trace to the instruction corpus itself. Data of this kind costs almost nothing to produce and needs no teacher model, so it is available to anyone who trains for multi-step reasoning.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Post-Training, Instruction Tuning, Training Data, LLMs, Synthetic Data, Reasoning

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_hbr7OpenReview ↗
Rating
1
Confidence
3(+1)
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an LLM mid-training method where instead of instruction-tuning on generic natural language data, they train on synthetic deductions in propositional logic. Their method results in increased downstream performance on various benchmarks for a variety of models at differing scales.

Critical strengths
  1. The method is simple and easy to understand. The fact that a simple intervention can yield non-trivial improvements to performance is noteworthy.
  2. The number of benchmarks, model families, and scales evaluated are numerous and makes progress toward supporting the claim that this approach can be used as a general mid-training method.
Critical weaknesses
  1. The paper's presentation is poor. The various experiments, particularly the numerous ablations, need to be motivated better so that the scientific questions they are answering are clearer. For example, it is unclear what the purpose of the 'shuffled lines' and 'derivations alone' ablations are as results are presented with little to no exposition.

  2. The choice of baseline mid-training data content and scale is unmotivated and lacks consideration. Comparisons against baselines trained on 100k examples randomly sampled from general instruction-tuning corpora do not sufficiently control for task specific over-fitting of the proposed method. It is possible that the downstream benchmarks share superficial similarities with the injected synthetic data.

Questions for authors
  1. Can the authors provide evidence that the proposed method does not simply fine-tune LLMs for the specific tasks that they are being evaluated on? For example, does your proposed mid-training method yield improved performance on tasks that do not have similar syntactic structure?

  2. Can the authors provide stronger evidence that the method they propose does not harm the LLM's performance on non-reasoning tasks far removed from symbolic reasoning (e.g. language modelling, general question answering in non-technical domains). Otherwise, it is hard to view this method as something other than domain specific fine-tuning.

4 comments

We thank the reviewer for the careful review and for recognising the simplicity of the method and the range of models, scales and benchmarks. Whether the gains come from fine-tuning on data that resembles the benchmarks is central to our claim as well. To test it, we trained four content ablations for two derivation mixtures (eight new models) and evaluated all models on further benchmarks far from symbolic reasoning.

The choice of baseline mid-training data content and scale is unmotivated [...] It is possible that the downstream benchmarks share superficial similarities with the injected synthetic data.

Can the authors provide evidence that the proposed method does not simply fine-tune LLMs for the specific tasks that they are being evaluated on? [...]

The training problems share little surface form with the benchmarks. They are chains over constants c0 to c25 with colour, direction and tree words as predicates, contain no negation and ask which of four candidates holds. None of the benchmarks uses this format. Accuracy nevertheless rises on BIG-Bench Hard tasks posed as English stories, by 12.5 points on web of lies and 8.1 on object tracking. If the gains came from superficial similarity, the tasks closest in form to logic problems would be expected to gain most, yet LSAT-AR, LSAT-LR, LogiQA and MMLU formal logic change by -1.9 to +1.8 points. The tasks that improve share a requirement, namely carrying a truth value or a state along a chain of steps, where the baseline often loses an intermediate value (Section 6.2).

The content ablations replace the same rows with other content. Procedural traces of column addition, sorting, base conversion and run-length encoding contain no logic and nothing resembling ProofWriter, yet recover most of its gain (+8.5 against +11.5 at 25 percent). This part of the gain therefore comes from step-by-step traces in general and cannot be explained by similarity to the benchmark. Two gains require the deduction content, on held-out generator problems (+3.8 and +9.5 against +0.2 and -0.1) and about half of the web of lies gain (+12.5 against +5.9).

The baseline is chosen to answer one question, namely what happens when part of a fixed instruction-tuning budget is spent on generated deductions. All mixtures therefore have the same 100,000 examples and one epoch, so that differences can be attributed to the content of the replaced rows. Appendix I varies the size from 25,000 to 200,000 examples. Section 5.3 repeats the comparison on Tulu 3, whose baseline answers "false" on 83 percent of ProofWriter items against 73 percent "true" for Dolci. The derivation procedure is learned on both corpora (66.4 percent of held-out derivations accepted on Tulu 3 against 0.0 for its baseline), while the ProofWriter and FOLIO gains are smaller on Tulu 3, whose baseline is already 14 points higher on ProofWriter.

Can the authors provide stronger evidence that the method they propose does not harm the LLM's performance on non-reasoning tasks [...]

We evaluated language modelling and question answering far from symbolic reasoning (Appendix K). On BoolQ, SciQ, TriviaQA, Natural Questions, LAMBADA, HellaSwag, PIQA and WinoGrande the derivation mixtures stay within 0.9 points of the baseline, and WikiText word perplexity changes by less than 0.05. No general benchmark falls by 4 points or more. The largest drop, 3.9 points on HumanEval at 25 percent, also appears under every content ablation, so we attribute it mainly to the code examples removed with the replaced rows.

The paper's presentation is poor. [...] it is unclear what the purpose of the 'shuffled lines' and 'derivations alone' ablations are [...]

We revised the presentation throughout, and all changes in the revised PDF are marked in blue. A new method section (Section 3, Algorithm 1) now describes the generator, the renderings, the checker and the mixing procedure, every number in the text was checked against the tables, and each ablation in Section 5.3 now begins with its purpose. Shuffled lines keep every token, the length and the conclusion of each derivation and destroy only the order of the steps. They test whether the model learns from the sequence of inferences or only from the facts and answers. The object-tracking gain disappears (+1.6 against +8.1) while the ProofWriter gain remains, so state tracking depends on the order of the steps and the ProofWriter gain mostly does not. Derivations alone (Appendix D.6) test whether instruction data is needed to apply the derivation skill to other tasks. Without instruction rows the model scores zero on every benchmark, so the transfer requires both.

We hope that this evidence addresses the reviewer's concerns. Should further evidence be needed for the reviewer to reconsider the assessment, we would be grateful to learn which experiments would be most informative, so that we can run them during the discussion period.

As the discussion period ends soon, we would be grateful to hear whether our responses and the revised PDF (changes marked in blue) address the reviewers' concerns, and which points remain open, so that we can still respond to them.

Thank you for your response. I now have a better understanding of the details pertaining to the issues that I raised in my original comment, and I raise my confidence.

Thank you for your reply. We noticed that the confidence was raised to 3 while the rating remained at 1, the lowest rating on the scale, although in our rebuttal we addressed all the points you raised. Since your comment suggests that the concerns have been clarified, we would like to ask whether the rating is intended as it stands, or whether it should also be updated. If any point is still open, we would be grateful to know which one, so that we can address it before the end of the discussion period.

Reviewer_Kf3EOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether programmatically generated symbolic deduction data can improve the reasoning capabilities of large language models. The authors construct synthetic multi-hop deduction problems and verifiable reasoning traces using a symbolic generator, without relying on an external teacher model. The generated data is represented in either formal symbolic notation or natural-language explanations and replaces a fraction of a general instruction-tuning dataset. The authors evaluate the approach on models from the Qwen2.5, Llama, and OLMo families, covering benchmarks such as ProofWriter, FOLIO, BIG-Bench Hard, and GPQA-Diamond. The experiments report improvements on several reasoning tasks and suggest that synthetic deduction data may help models maintain intermediate states during reasoning while also modifying answer-label biases introduced by instruction tuning.

Critical strengths

The paper explores an inexpensive and scalable alternative to model-generated synthetic data. By constructing deduction problems and reasoning traces programmatically, the approach avoids reliance on external teacher models and provides logically structured supervision. This is a practically relevant direction for improving the efficiency of LLM post-training.

Besides, the paper reports improvements on several reasoning benchmarks and evaluates multiple model families. The additional analyses of intermediate-state tracking, answer-label distributions, and different reasoning representations may provide useful insights into the effects of synthetic deduction supervision.

Critical weaknesses

The paper lacks a dedicated Methodology section, with key details of synthetic data generation, reasoning trace validation, and data replacement scattered across the experimental setup and appendices. This makes the proposed approach difficult to follow and reproduce. There are also inconsistencies between the text and tables, including conflicting replacement ratios and inaccurate descriptions of results.

The reported average improvement of 7.1 points across 13 metrics may be misleading, as several metrics overlap or belong to the same task families (e.g., ProofWriter, BBH, and GPQA). Moreover, Table 6 shows that substantial gains remain even after shuffling reasoning steps, while Table 8 suggests that improvements are concentrated in certain answer categories. These findings make it difficult to determine whether the gains come from better multi-step reasoning or other effects, such as answer calibration and synthetic data exposure.

Questions for authors
  1. Could the authors provide a clearer description of the full methodology, including deduction problem generation, reasoning trace verification, and how synthetic data is incorporated into instruction tuning?
  2. Why is the average improvement calculated across 13 overlapping metrics? Would a task-level average provide a more balanced evaluation?
  3. How can the authors distinguish genuine reasoning improvements from changes in answer-label preferences? In particular, why do shuffling reasoning steps in Table 6 still retain substantial gains?
  4. Table 3 states that both representations use a 10% replacement ratio, whereas Section 3.2 specifies 5% for Formal and 10% for English. The Qwen2.5-7B results in Table 3 also appear to match these respective settings in Table 1.
4 comments

We thank the reviewer for the detailed review and for checking the text against the tables, which helped us find and correct errors. To separate reasoning from changes in the answer prior, we trained four content ablations for two derivation mixtures and report per-class recall for every model.

The paper lacks a dedicated Methodology section [...]

Could the authors provide a clearer description of the full methodology, including deduction problem generation, reasoning trace verification, and how synthetic data is incorporated into instruction tuning?

The generator builds ground deductions over Horn clauses with abstract constants, adds distractor rules at every step and keeps only problems in which exactly one of four candidates is derivable, which we verified with a separate solver. Each proof is rendered either in a formal notation with numbered, justified lines or as English sentences with the same content. A forward-chaining checker accepts a derivation if every line is licensed by the premises or earlier lines and the last line is the correct answer; it is used only to score held-out problems. To build a mixture, a uniformly sampled share of the 100,000-example instruction corpus is replaced with derivations, and the model is fine-tuned for one epoch. This description is now given in a dedicated method section (Section 3) with Algorithm 1, and Section 4 contains the training and evaluation setup.

There are also inconsistencies between the text and tables [...]

Table 3 states that both representations use a 10% replacement ratio, whereas Section 3.2 specifies 5% for Formal and 10% for English. [...]

The caption of this table was incorrect, since the family comparison uses 5 percent formal and 10 percent English, as stated in the text and as shown by the Qwen2.5-7B row. The table (now Table 2) has a corrected caption, and every number in the text has been checked against the tables.

The reported average improvement of 7.1 points across 13 metrics may be misleading [...]

Why is the average improvement calculated across 13 overlapping metrics? Would a task-level average provide a more balanced evaluation?

We agree that a task-level average is more balanced. Of the 13 metrics, four were ProofWriter variants, three BIG-Bench Hard and three multi-hop QA, so ProofWriter received four times the weight of FOLIO. The average in Table 1 and Figure 1 is now taken first within each of five families (ProofWriter, FOLIO, BIG-Bench Hard, GPQA-Diamond and multi-hop QA) and then over the families. It is +6.5 points for formal derivations at 25 percent and +4.2 for English at 10 percent.

Moreover, Table 6 shows that substantial gains remain even after shuffling reasoning steps, while Table 8 suggests that improvements are concentrated in certain answer categories. [...]

How can the authors distinguish genuine reasoning improvements from changes in answer-label preferences? In particular, why do shuffling reasoning steps in Table 6 still retain substantial gains?

Part of the ProofWriter gain is a change in the answer prior, the Dolci baseline answers "true" on 73 percent of items against 40 percent gold. A shift of this prior alone does not raise balanced accuracy (mean recall over the gold labels), since the baseline says "true" about as often on false items (76 percent at depth 3) as on true items (73). At depth 3 balanced accuracy rises from 33.5 to 41.2 (formal, 5 percent), 43.1 (formal, 25 percent) and 43.9 (English, 10 percent) (Table 21). Recall of "false" rises from 21 to between 54 and 56, recall of "true" stays between 66 and 73, and recall of "unknown" stays at or below 7.

In the content ablations of formal derivations at 25 percent, answers alone give +4.4 on ProofWriter against +11.5 (balanced accuracy 37.0 against 43.1). Procedural traces without logic give +8.5 (42.9) and +3.9 against +4.3 on FOLIO, so here much of the gain beyond the prior comes from step-by-step traces in general. Two gains require the deduction content: on held-out generator problems the derivations give +3.8 (formal) and +9.5 (English) against +0.2 and -0.1 for procedural traces, and on web of lies +12.5 against +5.9.

Shuffled lines keep every token, the length and the conclusion, and the ProofWriter, FOLIO and web of lies gains are kept. The object-tracking gain is lost with both renderings (+1.6 and +1.7 against +8.1 and +6.5), as is about two thirds of the held-out gain with English (+3.3 against +9.5). The order of the lines therefore matters for state tracking and for English derivations, and little for ProofWriter. Sections 5.3 and 6.1 now discuss both effects, and the low recall of "unknown" is listed as a limitation.

We hope that these answers address the reviewer's concerns. Should further evidence be needed for the reviewer to reconsider the assessment, we would be grateful to learn which experiments would be most informative, so that we can run them during the discussion period.

As the discussion period ends soon, we would be grateful to hear whether our responses and the revised PDF (changes marked in blue) address the reviewers' concerns, and which points remain open, so that we can still respond to them.

Thank you to the authors for their detailed rebuttal. I have carefully reviewed their responses and most of my concerns have been addressed. I decided to raise my score to 3 (Weak Accept).

Thank you for reading our responses and for raising the score. You mention that most of the concerns have been addressed. We would be grateful to know which points remain open, so that we can address them in the remaining discussion period or in the camera-ready version.

Reviewer_r5ZFOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The author investigate whether program-generated deduction data containing abstract constants and no world knowledge can improve general instruction tuning. The authors replace a part (up to 25%) of a fixed-size instruction dataset with formal or English renderings of generated deductions for post-training. They report improvements on some benchmarks, such as ProofWriter, FOLIO, selected BIG-Bench Hard tasks, and GPQA-Diamond. Their analyses identify improved intermediate-state tracking and correction of answer-label biases as contributing mechanisms. The results suggest that procedural supervision can address specific reasoning deficiencies, although its effectiveness depends on the model and the rendering of the training data.

Critical strengths
  1. A useful investigation of program-generated data for post-training. This work provides deduction problems and checkable solutions without requiring a teacher model to generate each example. It offers a scalable, controlled source of supervision and investigates whether procedural structure alone can improve downstream reasoning.

  2. Systematic experiments with informative behavioral analyses. The study examines different mixture proportions, formal and English renderings, multiple model families, and several ablations. Holding the number of training examples fixed makes the replacement comparison informative. The analyses of intermediate-state tracking and answer-label bias also help explain when and why the additional data is useful.

Critical weaknesses
  1. Limited availability of reproducibility artifacts. Although the manuscript describes the generator and training procedure, the supplied materials do not provide access to the implementation, generated training datasets, or training and evaluation scripts. This limits independent reproduction and auditing of the data construction, verification procedures, and reported improvements. Releasing these artifacts with the exact experimental configurations would strengthen confidence in the results.

  2. Uncertain breadth of transfer and capability preservation. Many improvements occur on tasks involving logical inference or sequential state tracking, which resemble the reasoning structures in the generated data. The evidence therefore supports targeted skill transfer more clearly than a broad improvement in reasoning. The evaluated benchmarks provide useful checks, but some coding results remain missing, and broader capability preservation is not established.

Questions for authors
  1. Which structural properties do the generated deductions share with the benchmarks showing improvements? What evidence distinguishes targeted training of those shared skills from a broader improvement in reasoning ability?
  2. How should the claim that general benchmark performance remains unchanged be scoped, given the incomplete coding results and model-dependent negative effects? Does the current evidence support expectations about performance in settings evaluated by SWE-bench or τ2\tau^2-bench, or are those outside the intended scope?
  3. Under what conditions would you expect this approach to benefit models trained for specialized domains? Can the proposed behavioral analyses identify relevant reasoning deficiencies and predict whether deduction-data mixing would help before fine-tuning?
7 comments

We thank the reviewer for the thorough assessment of our work and for the constructive questions. To answer them, we ran new evaluations on coding, knowledge and language modelling benchmarks and trained Qwen2.5-Math-7B and Qwen2.5-Coder-7B.

Limited availability of reproducibility artifacts. [...]

The generator, checker, generated training mixtures and evaluation splits, training configs, evaluation tasks and analysis scripts are available at https://anonymous.4open.science/r/Lead-into-Gold-CB8F/. One script regenerates every mixture and evaluation split from fixed seeds, and the README lists the training and evaluation commands for each table. The paper now contains a reproducibility statement.

Uncertain breadth of transfer and capability preservation. [...]

How should the claim that general benchmark performance remains unchanged be scoped [...]? Does the current evidence support expectations about performance in settings evaluated by SWE-bench or tau-bench, or are those outside the intended scope?

With formal derivations at 25 percent HumanEval falls by 3.9 points, the largest drop among the general benchmarks we evaluated. With English derivations at 10 percent it rises by 2.0, and MBPP changes by at most 0.8 points for every mixture. The other drops are smaller, IFEval -2.2 and MATH500 -0.7 for formal at 25 percent and IFEval -0.9 and MATH500 -2.1 for English at 10 percent. BIG-Bench Hard rises by 1.7 and 2.0.

We attribute the HumanEval drop mainly to the code examples that are removed together with the replaced rows. These rows are sampled uniformly, so at 25 percent a quarter of the code examples in the instruction mixture is removed. Every content ablation at 25 percent loses HumanEval accuracy as well, including those without derivations (-2.8 for answers only, -3.3 for procedural traces, -1.8 for further Dolci reasoning examples). We did not count the code examples among the replaced rows directly. On Qwen2.5-Coder-7B the cost is larger, with HumanEval falling by 4.0 (formal) and 6.1 (English) and MBPP by at most 1.1 (Appendix M).

On BoolQ, SciQ, TriviaQA, Natural Questions, LAMBADA, HellaSwag, PIQA and WinoGrande the derivation mixtures stay within 0.9 points of the baseline, and WikiText word perplexity changes by less than 0.05 (Table 24). The claim in the paper now reads that no general benchmark falls by 4 points or more, and Appendix K lists every change (Tables 23 and 24). SWE-bench and tau-bench require multi-turn tool use, which single-turn instruction tuning does not train. We consider them outside the scope of this paper, and Section 7 lists them as future work.

Which structural properties do the generated deductions share with the benchmarks showing improvements? What evidence distinguishes targeted training of those shared skills from a broader improvement in reasoning ability?

The tasks that improve require carrying a truth value or a state along a chain of steps, as in ProofWriter, FOLIO, web of lies and object tracking. On these tasks the baseline often loses an intermediate value inside the chain, for example by reading the negation of its own previous step, whereas the trained models keep it (Section 6.2, Figure 2). Logic-shaped tasks without such a chain change little. LSAT-AR, LSAT-LR, LogiQA, MuSR and MMLU formal logic change by -1.9 to +1.8 points (Table 24), and multi-hop QA by -0.9 to +3.1 F1 (Appendix D.5). We agree that the evidence supports targeted transfer, and the introduction now describes it in these terms.

Under what conditions would you expect this approach to benefit models trained for specialized domains? Can the proposed behavioral analyses identify relevant reasoning deficiencies and predict whether deduction-data mixing would help before fine-tuning?

We trained Qwen2.5-Math-7B and Qwen2.5-Coder-7B (Appendix M) and stated before training the prediction that accuracy rises when the answer distribution moves toward the gold labels. The prediction held for both Coder mixtures and for Math with English, and it failed for Math with formal derivations (Table 27). The Coder baseline answers “true” on 62.5 percent of ProofWriter items against 40 percent gold, and the formal mixture removes this bias and gains 14.9 points. Under the formal mixture the distribution of the Math model moves away from the gold labels, while its accuracy still rises by 3.3. Over 8 combinations of model and mixture, the change in the answer distribution correlates with the gain (r = -0.74), but the bias measured before tuning does not predict it (r = 0.18). A probe of the answer prior before tuning is therefore not sufficient on its own to predict the benefit.

We hope that these answers address the reviewer’s concerns. Should further evidence be needed for the reviewer to reconsider the assessment, we would be grateful to learn which experiments would be most informative, so that we can run them during the discussion period.

As the discussion period ends soon, we would be grateful to hear whether our responses and the revised PDF (changes marked in blue) address the reviewers' concerns, and which points remain open, so that we can still respond to them.

Thank you for sharing the repository and running the additional evaluations. This release substantially address my concern about artifact access, eventhough it need more time to be reviewed as it is not released before the paper submission due. The new experiments strengthen the case for targeted transfer to tasks that require tracking intermediate states, and I agree that multi-turn tool-use benchmarks are outside the scope of this study. The HumanEval drop of up to 6.1 percentage points on Qwen2.5-Coder-7B is a meaningful trade-off. I would encourage you to highlight it in the main text and limit the claim that no general benchmark falls by four points or more to the Qwen2.5-7B experiments for which it holds.

Comments on the paper

2 comments

We thank all reviewers for their careful reading of the paper and for their constructive feedback, which has helped us improve it. To answer the reviews, we fine-tuned more than 30 additional models, among them four content ablations for two derivation mixtures and two specialist models, and evaluated all models on further benchmarks of coding, knowledge, reading comprehension and language modelling. The main changes are listed below, and each point is answered in the individual responses.

  1. A new method section (Section 3, Algorithm 1) describes the generator, the two renderings, the checker and the replacement of instruction rows (Kf3E, hbr7).
  2. The headline average in Table 1 and Figure 1 is computed over five task families, each counted once (Kf3E).
  3. Content ablations replace the same rows of the instruction mixture with answers only, derivations with shuffled lines, procedural traces without logic, or further Dolci reasoning examples (Section 5.3, Appendix L) (Kf3E, hbr7).
  4. Per-class recall and balanced accuracy are reported (Section 6.1), and the answer-prior account is tested on Qwen2.5-Math-7B and Qwen2.5-Coder-7B (Appendix M) (Kf3E, r5ZF).
  5. Further evaluations cover coding, knowledge, reading comprehension and language modelling (Appendix K) (r5ZF, hbr7).
  6. Code, generator, data-generation scripts, configs and evaluation scripts are available at https://anonymous.4open.science/r/Lead-into-Gold-CB8F/ (r5ZF).

All changes in the revised PDF are marked in blue.

The content ablations led us to state the main claim more precisely. Procedural traces of column addition, sorting, base conversion and run-length encoding, which are generated by a program but contain no logic, recover most of the ProofWriter and FOLIO gain. Accuracy on held-out generator problems and about half of the web of lies gain require the deduction content. The abstract, introduction and contributions have been revised accordingly. We would be grateful if the reviewers let us know whether any further evidence would be informative, so that we can run the corresponding experiments during the discussion period.

Brief update: we have uploaded a version of the PDF in which all changes to the manuscript are marked in blue, so that the revisions can be checked at a glance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.