Pith.

Software Engineering · arXiv:2610.18755 v1 · submitted 13 Oct 2026 · 87 references

PreciseCoder: Faithful and Generalizable Debugging Beyond Unit-Test Rewards

Yunkai Zhan, Miaosen Chai, Shangshang Wang, Jike Zhong, Yuqing Yang, Deqing Fu, Serina Chang, Wang Bill Zhu

The paper says

“Reinforcement learning (RL) from unit tests is the common recipe for post-training code models, but a unit test only asks whether the final program passes.”

more of the abstract

We analyze what this unit-test-only reward teaches a model during debugging and find that, while it raises pass rates, it also teaches the model to edit more and to hallucinate bugs in code that is already correct. We introduce PreciseCoder, a training algorithm that augments the unit-test reward with a signal for semantically verified, localized edits, which correlates with smaller edits and higher pass rates. On 4B, 9B, and 27B models, PreciseCoder reduces the average number of edit lines from 7-15to 3 at a pass rate on par with unit-test-only training. Trained only on single-line bugs, PreciseCoder can generalize to repository-level program repair, where it resolves more issues with smaller patches. Its precision also transfers from weaker to stronger models: given the reasoning trace of a 9B PreciseCoder model, frontier models such as GPT-5.6 Sol and Qwen3.8-Max gain 4.6-12.4 points in edit precision with no additional training. Linguistic analysis of the thinking traces shows that our reward encourages the model to reason about individual lines, name the buggy ones, and form consistent bug hypotheses.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 14 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Unit-test-only RL is plausibly a genuine cause of over-editing and bug hallucination, and the F1-augmented reward fix mostly holds; what is missing is the specification and code for the semantic verification machinery. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 87 references say.

The paper says, at Section 3.2, Eq. (1)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The reward's precision and recall depend on two procedures that the paper does not specify: aligning each predicted edit block to a reference fault block, and computing ess_i, the smallest subset of the predicted block that still repairs fault i by counterfactual test execution.

more of objection 1

Section 3.2 attributes the construction to Zhu et al. (2026b), but here it is used as a dense per-rollout training signal rather than an offline evaluation metric, and the central pass-rate-parity claim relies on this signal not degenerating into a noisy edit-size penalty. Appendix E reports only 5-7% wall-clock overhead, which is hard to reconcile with an exact search over the observed 7-15-line edits; if the implementation uses a greedy or single-line approximation, precision credit can be inflated for bloated blocks. Please specify the alignment and subset-search algorithms, or provide an ablation comparing exact and approximate variants, before the F1 term can be assessed. I am not asking for a re-derivation of the cited evaluation metric; the issue is the dense-training use and the overhead claim.

The paper says, at Section 4.1, Table 2

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

Table 11 reports a single seed for Qwen3.6-27B, and Table 2 gives no variance for the 27B rows. Because the pass-rate-parity claim at that size rests on a 3.0-point difference, please report multiple seeds or a confidence interval, or explicitly limit the parity claim to 4B and 9B.

The paper says, at Section 5.2, Tables 4 and 8

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The weak-to-strong transfer claim is a headline contribution, but Table 4 shows one donor seed and single-sample evaluation, while four of the five receivers' pass-rate deltas are small relative to likely sampling noise. Since Table 8 already reports the three-seed donor mean and standard deviation, please report those quantities in the main table so the reader can judge whether the pass-rate preservation and precision gains are stable across donor seeds.

5 minor objections

The paper says, at Appendix B.3

Pith says: minor objection 1 of 5

The lambda sweep is run under GRPO, while the main training runs use GDPO; please clarify whether the sweep was repeated under GDPO or whether the ordering argument in Appendix A.1 is intended to cover the actual algorithm.

The paper says, at Figure 1b

Pith says: minor objection 2 of 5

The caption labels 'single run / mean of 2 runs' but the legend does not identify which curve is which; please make the caption self-contained.

show 3 more

The paper says, at Figure 5

Pith says: minor objection 3 of 5

The SWE-bench resolved-rate differences are small and no confidence intervals are reported; please add intervals or per-instance variance so the reader can judge the 25.9-32.7 and 67.9-70.8 ranges.

The paper says, at Appendix C.2

Pith says: minor objection 4 of 5

The RECAP refiner is trained with reasoning traces generated by DeepSeek V4.1 Flash, which introduces a stronger teacher into a baseline; please state explicitly why this does not give RECAP an advantage relative to the other baselines.

The paper says, at Section 3.2

Pith says: minor objection 5 of 5

The sentence 'a passing program has by definition resolved every injected fault' assumes the unit tests cover every injected fault; please state this assumption explicitly, since the gated recall definition depends on it.

What would settle it

Re-implement Eq. (1) exactly from the released code or the PDB implementation: write out the explicit alignment rule and replace the subset search with exhaustive enumeration over blocks of size ≤ 4 (≤ 16 counterfactual test executions per block, feasible on a sample of groups), then replay the three PreciseCoder seeds' training rollouts and evaluation outputs through both the reported and the exact F1. If the per-group rank correlation is below ~0.95, or the within-group advantage ordering (Appendix A.2 collapse analysis) flips on more than ~5% of groups, the F1 signal is materially approximate and Table 2's pass-rate parity must be re-verified by retraining at 4B with the exact signal. Report the value of ε in min(|ess_i|, |δ⋆_i| + ε) as part of the check, since it caps per-fault precision credit.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 14 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 5 min 45 s · 31,570 tokens read, 17,696 written
  3. Editor wrote the verdict above 35 s · 32,798 tokens read, 4,723 written
  4. Referee major revision: 3 major objections, 5 minor 59 s · 32,001 tokens read, 8,189 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Worth a careful read. This is the strongest empirical case I have seen for a diagnosis that has been floating around: unit-test-only RL is what makes code models over-edit and invent bugs in correct code. The evidence is the GRPO control on bug-free programs (no-op rate drops from 31–37% base to 19–31% under the unit-test reward), and then the F1-augmented reward bringing it back up to 45–52% while pass rate stays on par at all three sizes (55.3/62.5/79.3 vs 55.1/63.9/76.3) and mean edits fall from roughly 8–14 lines to about 3. That is a direct intervention on the reward, not a curve fit.

What is new: the diagnosis itself, plus the mechanism. Prior work penalizes edit size and visibly pays for it in pass rate (RECAP and EA-GRPO at 4B: 38.1 and 53.3 vs 55.1); the semantically verified F1 reward keeps pass rate while improving precision, and the Table 3 ablations show less faithful signals lose 1.8–5 points, so the signal is doing something specific. The weak-to-strong trace transfer in Table 4 has a real control (the untrained 9B trace costs the 27B receiver 7.4 pass points), and the largest precision gains go to the stronger receivers. The SWE-bench and PDB-Wild numbers are checked by external tests and gold patches, so the circularity burden is low.

Soft spots: Eq. (1)'s machinery—block alignment and the smallest-subset counterfactual search—is not specified here. The paper invokes the PDB paper and promises code, and since the same machinery computes both the training reward and the reported precision on PDB Test and PDB-Wild, an approximate search could inflate both together. That is the one place I want shipped code before fully trusting the precision numbers. The stress-test's overhead argument does not move me much: 5–7% relative to unit-test execution, with single-line reference faults in training, is plausible even for a non-exact search, so I would not sink the paper on that. Minor issues: the 27B row and all generalization and agentic runs are single-seed, and “root cause” in Section 2 is stronger than the correlational training-dynamics evidence warrants. Both are fixable.

Net: the diagnosis is well-evidenced, the fix is different in mechanism from edit-size penalties, and the generalization results are externally validated. I would bring this to a code-RL reading group and, if it were my venue, send it to serious referees.

The strongest objection

The central claim has three pillars: pass-rate parity with unit-test-only GRPO, a roughly 3-line average edit, and 24–35-point precision gains. The first two are machinery-independent (unit tests and line-level diffs), but the mechanism that supposedly makes parity and fidelity compatible — the semantically verified F1 of Eq. (1) — is not specified. Computing rec and prec requires two procedures this paper does not give: (i) aligning each predicted block δ̂_i to a reference fault block δ⋆_i, and (ii) finding ess_i, "the smallest subset of the predicted block that still repairs fault i", by counterfactual test execution. Section 3.2 credits the construction to the authors' PDB benchmark paper (Zhu et al., 2026b), where it serves offline evaluation, not a dense per-rollout training signal. Two failure modes make this load-bearing. First, over-editing rollouts are large (GRPO's mean edits are 11.9–13.8 lines against single-line reference faults), so if alignment is positional or greedy, a correct fix can be attached to the wrong reference block; the counterfactual swap then fails, fix_i = 0, and a valid repair earns no credit. Second, an exact smallest-subset search over a k-line block costs up to 2^k test executions, yet Appendix E reports only 5–7% wall-clock overhead — consistent with greedy line deletion or single-line candidates, which would inflate precision credit for bloated blocks and thereby partially reward over-editing. The same machinery computes the reported precision on PDB Test and PDB-Wild, so an approximation would inflate the training signal and the headline metric together. Table 3 provides the sensitivity check: less faithful signals (unit·prec, ungated F1, EA-GRPO's size penalty) lose 1.8–5 pass points, so parity is exactly where an unfaithful F1 would surface. The gold-patch dependency is acknowledged in Appendix E; the alignment/subset-search dependency is not.

The central claim is that a semantically verified F1 reward is what lets PreciseCoder cut edits from 7–15 lines to 3 while keeping pass rate on par with unit-test-only GRPO — a combination that, per the paper's own Table 3, no unfaithful reward achieves. For that claim, everything hangs on the F1 signal computed per rollout during training: it must actually measure whether each edited line is necessary. Eq. (1) defines the quantities (fix_i via counterfactual swap, ess_i as the smallest repairing subset), but the two procedures that make them computable — block-to-fault alignment and the subset search — are covered only by a deferral to the authors' PDB benchmark paper, where the machinery serves offline evaluation rather than a dense per-rollout reward. The claimed 5–7% overhead (Appendix E) is evidence of approximation, or of implausibly small search spaces given 11.9–13.8-line mean edits. I find no internal contradiction in the normalization story (Appendix A.1's λ < 0.37 bound with λ = 1/3 is consistent), and the controls are strong: three seeds at 4B and 9B, the reward-weight sweep in Appendix B.3, and the weak-to-strong transfer with all-9B-donor controls in Table 8. The 27B single-seed runs and the "root cause" phrasing in Section 2 are real but lesser issues, and the gold-patch dependency is explicitly acknowledged in Appendix E. One deepening over the reader's account: the same unverified machinery computes the reported precision on PDB Test and PDB-Wild, so a systematic approximation would inflate both the training signal and the headline precision numbers; pass rate and edited lines would survive, but the paper's distinctive mechanism claim would not. The concern is concrete and testable — specify the alignment rule, release the implementation, and compare against an exact re-computation — so it strengthens the CONDITIONAL verdict rather than forcing rejection. Verdict unchanged.

The referee's report

If the results hold, the paper makes a solid contribution: it turns a known failure mode (test-passing but over-edited patches) into a trainable reward and validates it with external metrics rather than its own scoring function. Strengths include the clean-rows control, OOD benchmarks, lambda sweep, multiple seeds at 4B/9B, explicit appendix limitations, and the use of unit-test pass rates and line-level diffs for the headline claims. The main risk is not circularity but the under-specified semantic-verification machinery behind Eq. (1).

Pith's reading of the paper (a machine's summary, not the authors')

Reinforcement learning from unit tests is the standard way to post-train code models, but this paper argues the reward is looking at the wrong thing: a unit test only scores the final program, so training teaches the model to pass tests by rewriting more code and even to invent bugs in programs that are already correct. The paper proposes PreciseCoder, which adds a second reward term, an F1 score of edit precision and recall verified semantically fault by fault against the reference repair. Across 4B, 9B and 27B models, the combined reward keeps the pass rate on par with unit-test-only training while cutting the average edit from 7 to 15 lines down to about 3 and raising edit precision by 24 to 29 points, and it raises the no-op rate on correct code from 31 to 37 percent up to 47 to 52 percent even though training never shows a correct program. Trained only on single-line bugs, the behavior carries to multi-line, repository-level and agentic repair, and the reasoning trace of a 9B model transfers precision to stronger models it cannot match on correctness. The sympathetic reading: over-editing is not the inevitable price of reinforcement learning but a reward-design artifact, curable by making the reward verify the edit rather than only the outcome.

What carries the argument. The load-bearing object is the reward $r_\lambda = (1-\lambda)\cdot\mathrm{unit} + \lambda\cdot F_1$, where the $F_1$ term is the harmonic mean of two semantically verified scores. Recall asks which reference faults the predicted edit actually repairs, gated on the unit test; precision aligns each predicted edit block to a reference fault block and credits only the smallest subset of the predicted block that still repairs that fault under counterfactual test execution (Eq. 1). That counterfactual step is what makes the reward see the edit: it converts a preference for small edits into a check of necessity, which is why precision correlates with small edits (above 0.81) without losing the unit-test correlation that recall carries (above 0.71). The second piece of machinery is GDPO's decoupled normalization, which standardizes the two terms within each group before mixing so the weight $\lambda$ cannot collapse unevenly across groups; under it, a passing completion keeps a larger advantage than a failing one whenever $\lambda < 2/(2+\sqrt{2G})$, and the paper uses $\lambda = 1/3$ at group size 6. A further emergent mechanism: on correct code the reward teaches the model to propose bug hypotheses and then reject them, and the no-op rate climbs with the number of hypotheses proposed and dismissed (from 21 to 68 percent between quartiles for PreciseCoder-9B), which is how restraint appears without ever training on a correct program.

Core claim. The paper's central claim has a diagnostic half and a constructive half. The diagnostic half: unit-test-only GRPO makes pass rate and edit size rise together, keeps rewarding over-editing after the pass rate plateaus, and lowers the no-op rate on bug-free validation code from 30.8 percent to 20.5 percent during training, so the reward itself teaches bug hallucination even though no training row contains a correct program. The constructive half: a reward $r_\lambda = (1-\lambda)\cdot\mathrm{unit} + \lambda\cdot F_1$, where $F_1$ is the harmonic mean of a test-gated recall and a semantically verified edit precision, is the only tested reward that lands in the desired region of the correlation analysis, correlating with both the unit-test outcome (above 0.62) and, among passing samples, with small edits (above 0.82). With per-term normalization and $\lambda = 1/3$, the trained policy keeps pass rate on par with unit-test-only GRPO at all three model sizes (55.3 vs 55.1 at 4B, 62.5 vs 63.9 at 9B, 79.3 vs 76.3 at 27B) while editing about three lines per task and lifting edit precision to frontier-model levels; on correct code, the no-op rate rises to 47 to 52 percent and damage falls below the base model. The same policy, trained only on single-line single-file bugs, resolves more SWE-Bench Verified issues than the base model or unit-test-only GRPO with smaller patches, and its think block, prefilled into five stronger receivers, raises their edit precision by 4.6 to 12.4 points with no weight access or further training.

The premise it rests on. Everything rests on the counterfactual verification in Eq. (1): each predicted edit block must be aligned to the right reference fault block, and the smallest subset that still repairs fault $i$ must be found by actually re-running the tests; if the alignment misattributes lines or the subset search is approximated away for cost, the precision term becomes noise and the F1 reward degenerates into an unreliable edit-size penalty, and separately every training task needs a reference repair to exist, a limitation the paper acknowledges.

If the paper is right (Pith's inferences, not the authors')

Unit-test-only RL is a diagnosable and curable cause of over-editing: switching to a semantically verified F1 reward keeps pass rate on par while cutting edits roughly fivefold, so pass rate and edit fidelity are not in inherent tension.

Edit-size penalties are the wrong lever: baselines that penalize changed lines (EA-GRPO, RECAP) trade 3 to 17 pass-rate points for smaller edits because edit size barely correlates with correctness ($r \approx 0.13$ to $0.14$), whereas the F1 reward correlates with both.

Precision learned on single-line, single-file bugs is not brittle: it transfers to multi-line and repository-level faults, to SWE-Bench Verified, and to five-turn agentic debugging, where a 9B model reaches 67.4 percent precision at a 56.3 percent pass rate.

Debugging discipline travels through the reasoning trace alone: prefilling PreciseCoder-9B's think block raises five stronger receivers' edit precision by 4.6 to 12.4 points, while an untrained donor's block costs the 27B receiver 7.4 pass-rate points, so the trace carries trained judgment, not mere length.

Naming the buggy line is the operative behavior: in every model examined, tasks whose thinking quotes every buggy line verbatim are solved with 4 to 19 points higher recall, and the share of tasks with full verbatim grounding rises from 36 to 43 percent up to 67 to 71 percent under PreciseCoder.

If the diagnostic holds, pass rate alone is a misleading reward and evaluation metric for code repair: over-editing patches that pass their tests still carry regression risk on behavior the tests do not cover, so agentic-coding benchmarks should routinely report edit-level fidelity.

The principle demonstrated here, credit an action only when it is verified necessary against a reference, is a general recipe for agentic reinforcement learning, not just debugging; the paper gestures at this for long-horizon tasks where unneeded actions are chances to break something, but leaves the construction of the reference open.

Replace the counterfactual subset search with a cheaper proxy, such as line coverage of the failing test, and measure how much of the precision gain and pass-rate parity survives; the paper reports full semantic verification costs only 5 to 7 percent extra wall-clock time, so a proxy would matter mainly for scaling to longer rollouts or larger groups.

The weak-to-strong result suggests the think block transmits a decision procedure, name the buggy line and weigh keeping each edit, rather than a style; one could ablate the block to keep only the quoted buggy lines and verdicts, stripped of prose, and check whether the precision transfer to receivers survives.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Reinforcement learning (RL) from unit tests is the common recipe for post-training code models, but a unit test only asks whether the final program passes. We analyze what this unit-test-only reward teaches a model during debugging and find that, while it raises pass rates, it also teaches the model to edit more and to hallucinate bugs in code that is already correct. We introduce PreciseCoder, a training algorithm that augments the unit-test reward with a signal for semantically verified, localized edits, which correlates with smaller edits and higher pass rates. On 4B, 9B, and 27B models, PreciseCoder reduces the average number of edit lines from 7-15to 3 at a pass rate on par with unit-test-only training. Trained only on single-line bugs, PreciseCoder can generalize to repository-level program repair, where it resolves more issues with smaller patches. Its precision also transfers from weaker to stronger models: given the reasoning trace of a 9B PreciseCoder model, frontier models such as GPT-5.6 Sol and Qwen3.8-Max gain 4.6-12.4 points in edit precision with no additional training. Linguistic analysis of the thinking traces shows that our reward encourages the model to reason about individual lines, name the buggy ones, and form consistent bug hypotheses.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-14

The authors' main claim, as the classifier picked it out 0.81

“We introduce PreciseCoder, a training algorithm that augments the unit-test reward with a signal for semantically verified, localized edits, which correlates with smaller edits and higher pass rates.”

How the main claim is stated: asserted 0.99

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.31 · human participants 0.03 · controlled comparison 0.93 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.95 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Reinforcement learning (RL) from unit tests is the common recipe for post-training code models, but a unit test only asks whether the final program passes. background 0.99 0.06 0.03 asserted 1.00
We analyze what this unit-test-only reward teaches a model during debugging and find that, while it raises pass rates, it also teaches the model to edit more and to hallucinate bugs in code that is already correct. result 0.89 0.95 0.22 asserted 1.00
We introduce PreciseCoder, a training algorithm that augments the unit-test reward with a signal for semantically verified, localized edits, which correlates with smaller edits and higher pass rates. method 0.55 0.92 0.39 asserted 0.99
On 4B, 9B, and 27B models, PreciseCoder reduces the average number of edit lines from 7-15to 3 at a pass rate on par with unit-test-only training. result 1.00 0.95 0.25 asserted 1.00
Trained only on single-line bugs, PreciseCoder can generalize to repository-level program repair, where it resolves more issues with smaller patches. result 1.00 0.93 0.18 asserted 0.97
Its precision also transfers from weaker to stronger models: given the reasoning trace of a 9B PreciseCoder model, frontier models such as GPT-5.6 Sol and Qwen3.8-Max gain 4.6-12.4 points in edit precision with no additional training. result 1.00 0.88 0.17 asserted 1.00
Linguistic analysis of the thinking traces shows that our reward encourages the model to reason about individual lines, name the buggy ones, and form consistent bug hypotheses. interpretation 0.50 0.94 0.09 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-14: 1 request, 5,822 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (14)
Figure 1 of the paper
Figure 1. (a) Illustration of over-editing and bug hallucination phenomenon; (b) Training dynamics
Figure 2 of the paper
Figure 2. Pearson r averaged over all training tasks (circle 4B, square 9B; n=16 samples per task). unit is constant when passing and is plotted at y=0. 3.2 HOW DIFFERENT REWARD TERMS CORRELATE Next, we analyze how candidate reward components correlate with one another on the Qwen3.5-4B and 9B base models (
Figure 3 of the paper
Figure 3. No-op rate on GT Debug scales lin￾early with precision
Figure 4 of the paper
Figure 4. On PDB-Wild multi-line and repository-level repairs, PreciseCoder matches or outper
Figure 5 of the paper
Figure 5. PreciseCoder resolves more issues with smaller
Figure 8 of the paper
Figure 8. GT Debug no-op rate by within-model quartile of bug hypotheses. More hypotheses, more restraint. rate rises from 21% to 68% between the quarter of tasks with the fewest hypotheses and the quarter with the most, whereas the 9B unit-test-only model stays near 30%, as shown in
Figure 9 of the paper
Figure 9. Reward-weight sweep on Qwen3.5-4B GRPO, PDB Test. Lines are seed means; whiskers
Figure 10 of the paper
Figure 10. Common word markers in the thinking traces.
Figure 11 of the paper
Figure 11. What each training changed in the 9B’s thinking on the same PDB Test task: unit-test
Figure 12 of the paper
Figure 12. PDB Test edit precision against pass rate (bubble area = edited lines). Arrows join each
Figure 13 of the paper
Figure 13. Chain-of-thought transfer on PDB Test. (a) A PreciseCoder think block prefilled into its
Figure 14 of the paper
Figure 14. GT Debug under the minimal-edit (circles) and abstain-or-fix (triangles) prompts. Up

Showing 12 of the paper's 14 figures. Figures are the authors', reproduced from the paper.

References (87)
  1. [1] Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Y. and Guo, Daya , journal =. 2024 , url = on Pith
  2. [2] Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and others , journal =. doi:10.1038/s41586-025-09422-z , year = on Pithdoi:10.1038/s41586-025-09422-z
  3. [3] Sheng, Guangming and Zhang, Chi and Ye, Zilingfeng and Wu, Xibin and Zhang, Wang and Zhang, Ru and Peng, Yanghua and Lin, Haibin and Wu, Chuan , booktitle =. doi:10.1145/3689031.3696075 , year = on Pithdoi:10.1145/3689031.3696075
  4. [4] Liu, Shih-Yang and Dong, Xin and Lu, Ximing and Diao, Shizhe and Belcak, Peter and Liu, Mingjie and Chen, Min-Hung and Yin, Hongxu and Wang, Yu-Chiang Frank and Cheng, Kwang-Ting and Choi, Yejin and Kautz, Jan and Molchanov, Pavlo , journal =. 2026 , note = on Pith
  5. [5] Nie, Wenhua and Wu, Jianan and Liu, Junlin and Li, Ziwei and Lin, Zheng and Zhang, Zijian and Fan, Yilong and Zheng, Haoran and Jang, Jyh-Shing Roger , journal =. 2026 , url = on Pith
  6. [6] Ichihara, Yuki and Jinnai, Yuu and Morimura, Tetsuro and Sakamoto, Mitsuki and Mitsuhashi, Ryota and Uchibe, Eiji , journal =. 2025 , note = on Pith
  7. [7] Bay, Yong Yi and Yearick, Kathleen A. , journal =. 2026 , url = on Pith
  8. [8] Liu, Zichen and Chen, Changyu and Li, Wenjun and Qi, Penghui and Pang, Tianyu and Du, Chao and Lee, Wee Sun and Lin, Min , journal =. 2025 , url = on Pith
  9. [9] Yu, Qiying and Zhang, Zheng and Zhu, Ruofei and Yuan, Yufeng and Zuo, Xiaochen and Yue, Yu and others , booktitle =. doi:10.52202/085713-3775 , year = on Pithdoi:10.52202/085713-3775
  10. [10] Wei, Yuxiang and Duchenne, Olivier and Copet, Jade and Carbonneaux, Quentin and Zhang, Lingming and Fried, Daniel and Synnaeve, Gabriel and Singh, Rishabh and Wang, Sida I. , booktitle =. doi:10.52202/085713-2629 , year = on Pithdoi:10.52202/085713-2629
  11. [11] Wang, Longwen and Liu, Yirui and Wu, Xuan'er and Hu, Xiaohui and Fan, Yuankai and Yu, Kaidong and Weng, Qizhen and Xi, Wei and Li, Xuelong , journal =. 2026 , url = on Pith
  12. [12] Li, Jia and Su, Yuxin and Peng, Ting and Huang, Hailiang and Deng, Yuetang and Lyu, Michael R. , journal =. 2026 , url = on Pith
  13. [13] Skalse, Joar and Hammond, Lewis and Griffin, Charlie and Abate, Alessandro , booktitle =. doi:10.24963/ijcai.2022/476 , year = on Pithdoi:10.24963/ijcai.2022/476
  14. [14] Yang, Boyang and Cai, Zijian and Jin, Shunfu and Tian, Haoye , journal =. 2026 , url = on Pith
  15. [15] Zhu, Wang Bill and Chai, Miaosen and Wang, Shangshang and Liu, Yejia and Bian, Song and Dong, Honghua and Neiswanger, Willie and Jia, Robin , booktitle =. 2026 , url = on Pith
  16. [16] Zhuo, Terry Yue and Vu, Minh Chien and Chim, Jenny and Hu, Han and Yu, Wenhao and Widyasari, Ratnadira and others , booktitle =. 2025 , url = on Pith
  17. [17] Jain, Naman and Han, King and Gu, Alex and Li, Wen-Ding and Yan, Fanjia and Zhang, Tianjun and Wang, Sida and Solar-Lezama, Armando and Sen, Koushik and Stoica, Ion , booktitle =. 2025 , url = on Pith
  18. [18] Mark Chen and Jerry Tworek and Heewoo Jun and Qiming Yuan and Henrique Ponde Oliveira Pinto and Jared Kaplan and Harri Edwards and Yuri Burda and Nicholas Joseph and Greg Brockman and others , journal =. 2021 , url = on Pith
  19. [19] Liu, Jiawei and Xia, Chunqiu Steven and Wang, Yuyao and Zhang, Lingming , booktitle =. 2023 , url = on Pith
  20. [20] Gu, Alex and Roziere, Baptiste and Leather, Hugh James and Solar-Lezama, Armando and Synnaeve, Gabriel and Wang, Sida , booktitle =. 2024 , url = on Pith
show 67 more
  1. [21] Tian, Runchu and Ye, Yining and Qin, Yujia and Cong, Xin and Lin, Yankai and Pan, Yinxu and Wu, Yesai and Haotian, Hui and Weichuan, Liu and Liu, Zhiyuan and Sun, Maosong , booktitle =. 2024 , address =. doi:10.18653/v1/2024.findings-acl.247 , url = on Pithdoi:10.18653/v1/2024.findings-acl.247
  2. [22] Jimenez, Carlos E. and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik , booktitle =. 2024 , url = on Pith
  3. [23] Yang, John and Lieret, Kilian and Jimenez, Carlos E. and Wettig, Alexander and Khandpur, Kabir and Zhang, Yanzhe and Hui, Binyuan and Press, Ofir and Schmidt, Ludwig and Yang, Diyi , booktitle =. 2025 , url = on Pith
  4. [24] Yang, John and Jimenez, Carlos E. and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik and Press, Ofir , booktitle =. 2024 , url = on Pith
  5. [25] Muennighoff, Niklas and Yang, Zitong and Shi, Weijia and Li, Xiang Lisa and Li, Fei-Fei and Hajishirzi, Hannaneh and Zettlemoyer, Luke and Liang, Percy and Candès, Emmanuel and Hashimoto, Tatsunori , journal =. 2025 , url = on Pith
  6. [26] Cheng, Xinyuan and Chen, Beiduo and Mondorf, Philipp and Plank, Barbara , journal =. 2026 , url = on Pith
  7. [27] Wu, Tong and Xiang, Chong and Wang, Jiachen T. and Suh, G. Edward and Mittal, Prateek , journal =. 2025 , url = on Pith
  8. [28] Yang, Wang and Yue, Xiang and Chaudhary, Vipin and Han, Xiaotian , journal =. 2025 , url = on Pith
  9. [29] Burns, Collin and Izmailov, Pavel and Kirchner, Jan Hendrik and Baker, Bowen and Gao, Leo and Aschenbrenner, Leopold and Chen, Yining and Ecoffet, Adrien and Joglekar, Manas and Leike, Jan and Sutskever, Ilya and Wu, Jeff , booktitle =. 2024 , url = on Pith
  10. [30] Liu, Alisa and Han, Xiaochuang and Wang, Yizhong and Tsvetkov, Yulia and Choi, Yejin and Smith, Noah A. , booktitle =. 2024 , url = on Pith
  11. [31] Ke, Changxin and Zhang, Rui and Guo, Jiaming and Wen, Yuanbo and Ding, Li and Wang, Shuo and Zhu, Xuyuan and Peng, Xiong and Huang, Di and Du, Zidong and Hu, Xing and Guo, Qi and Chen, Yunji , booktitle =. doi:10.18653/v1/2026.acl-long.1343 , url = on Pithdoi:10.18653/v1/2026.acl-long.1343
  12. [32] Qi, Zichao and Long, Fan and Achour, Sara and Rinard, Martin , booktitle =. doi:10.1145/2771783.2771791 , year = on Pithdoi:10.1145/2771783.2771791
  13. [33] Smith, Edward K. and Barr, Earl T. and Le Goues, Claire and Brun, Yuriy , booktitle =. doi:10.1145/2786805.2786825 , year = on Pithdoi:10.1145/2786805.2786825
  14. [34] Price, George R. , journal =. doi:10.1038/227520a0 , year = on Pithdoi:10.1038/227520a0
  15. [35] Ng, Andrew Y. and Harada, Daishi and Russell, Stuart , booktitle = on Pith
  16. [36] 2024 , month = aug, note = on Pith
  17. [37] Li, Yujia and Choi, David and Chung, Junyoung and Kushman, Nate and Schrittwieser, Julian and Leblond, R. Science , volume =. doi:10.1126/science.abq1158 , year = on Pithdoi:10.1126/science.abq1158
  18. [38] Rozi. arXiv preprint arXiv:2308.12950 , year = on PithPith's reviewarXiv
  19. [39] Guo, Daya and Zhu, Qihao and Yang, Dejian and Xie, Zhenda and Dong, Kai and Zhang, Wentao and Chen, Guanting and Bi, Xiao and others , journal =. 2024 , url = on Pith
  20. [40] Hui, Binyuan and Yang, Jian and Cui, Zeyu and Yang, Jiaxi and Liu, Dayiheng and Zhang, Lei and Liu, Tianyu and Zhang, Jiajun and others , journal =. 2024 , url = on Pith
  21. [41] Wang, Xingyao and Li, Boxuan and Song, Yufan and Xu, Frank F. and Tang, Xiangru and Zhuge, Mingchen and Pan, Jiayi and Song, Yueqi and others , booktitle =. 2025 , url = on Pith
  22. [42] Xia, Chunqiu Steven and Deng, Yinlin and Dunn, Soren and Zhang, Lingming , journal =. 2024 , url = on Pith
  23. [43] Pan, Jiayi and Wang, Xingyao and Neubig, Graham and Jaitly, Navdeep and Ji, Heng and Suhr, Alane and Zhang, Yizhe , booktitle =. 2025 , url = on Pith
  24. [44] Yang, Zonghan and Wang, Shengjie and Fu, Kelin and He, Wenyang and Xiong, Weimin and Liu, Yibo and Miao, Yibo and Gao, Bofei and others , journal =. 2025 , url = on Pith
  25. [45] Luo, Michael and Jain, Naman and Singh, Jaskirat and Tan, Sijun and Patel, Ameen and Wu, Qingyang and Ariyak, Alpay and Cai, Colin and others , howpublished =. 2025 , note = on Pith
  26. [46] Lambert, Nathan and Morrison, Jacob and Pyatkin, Valentina and Huang, Shengyi and Ivison, Hamish and Brahman, Faeze and Miranda, Lester James V. and Liu, Alisa and others , journal =. 2024 , url = on Pith
  27. [47] Chen, Zhaoling and Tang, Xiangru and Deng, Gangda and Wu, Fang and Wu, Jialong and Jiang, Zhiwei and Prasanna, Viktor and Cohan, Arman and Wang, Xingyao , journal =. 2025 , url = on Pith
  28. [48] Zhu, Tongyao and Lim, Wei Hern and Kan, Min-Yen , booktitle =. 2026 , url = on Pith
  29. [49] Dai, Zhenlong and Chen, Bingrui and Zhao, Zhuoluo and Tang, Xiu and Wu, Sai and Yao, Chang and Gao, Zhipeng and Chen, Jingyuan , booktitle =. 2025 , url = on Pith
  30. [50] Luo, Wenqiang and Keung, Jacky and Shi, Xiaoyu and Sun, Yicheng and Yang, Boyang and Yang, Zhou and Tian, Haoye , journal =. 2026 , url = on Pith
  31. [51] Ahmed, Toufique and Ganhotra, Jatin and Shinnar, Avraham and Hirzel, Martin , journal =. 2025 , url = on Pith
  32. [52] Tan, Shin Hwei and Yoshida, Hiroaki and Prasad, Mukul R. and Roychoudhury, Abhik , booktitle =. doi:10.1145/2950290.2950295 , year = on Pithdoi:10.1145/2950290.2950295
  33. [53] Watanabe, Miku and Li, Hao and Kashiwa, Yutaro and Reid, Brittany and Iida, Hajimu and Hassan, Ahmed E. , journal =. 2025 , url = on Pith
  34. [54] Singhal, Manav and Aggarwal, Tushar and Awasthi, Abhijeet and Natarajan, Nagarajan and Kanade, Aditya , booktitle =. 2024 , url = on Pith
  35. [55] Zhong, Ming and Zhou, Xiang and Chang, Ting-Yun and Wang, Qingze and Xu, Nan and Si, Xiance and Garrette, Dan and Upadhyay, Shyam and others , booktitle =. 2026 , url = on Pith
  36. [56] Gloaguen, Thibaud and M. arXiv preprint arXiv:2605.07769 , year = on PithPith's reviewarXiv
  37. [57] Liu, Xun and Zhang, Yi Evie and Kasprova, Vira and Rabbani, Parisa and Zahraei, Pardis Sadat and Zhang, Tianyu and Ebrahimpour-Boroojeny, Ali and Chandrasekaran, Varun , journal =. 2026 , url = on Pith
  38. [58] Qu, Yubin and Zhang, Ying and Zhang, Yanjun and Deng, Gelei and Li, Yuekang and Zhang, Leo Yu and Liu, Yi , journal =. 2026 , url = on Pith
  39. [59] McAleese, Nat and Pokorny, Rai Michael and Ceron Uribe, Juan Felipe and Nitishinskaya, Evgenia and Trebacz, Maja and Leike, Jan , journal =. 2024 , url = on Pith
  40. [60] Lu, Junyi and Jiang, Lili and Li, Xiaojia and Fang, Jianbing and Zhang, Fengjun and Yang, Li and Zuo, Chun , booktitle =. 2025 , url = on Pith
  41. [61] Huang, Jie and Chen, Xinyun and Mishra, Swaroop and Zheng, Huaixiu Steven and Yu, Adams Wei and Song, Xinying and Zhou, Denny , booktitle =. 2024 , url = on Pith
  42. [62] Olausson, Theo X. and Inala, Jeevana Priya and Wang, Chenglong and Gao, Jianfeng and Solar-Lezama, Armando , booktitle =. 2024 , url = on Pith
  43. [63] Bi, Ziqian and Wang, Yinzhi and Wang, Tianyang and Hao, Junfeng and Peng, Benji and Weng, Wenqian and Gu, Jiayi and Pang, Jacqueline and Song, Xinyuan , journal =. 2025 , url = on Pith
  44. [64] Asawa, Parth and Zhu, Alan and O'Neill, Abigail and Zaharia, Matei and Dimakis, Alexandros G. and Gonzalez, Joseph E. , booktitle =. 2026 , url = on Pith
  45. [65] Lingam, Vijay and Golatkar, Aditya and Pal, Anwesan and Vo, Ben and Sadagopan, Narayanan and Achille, Alessandro and Huan, Jun and Deoras, Anoop and Soatto, Stefano , booktitle =. 2026 , url = on Pith
  46. [66] Yang, Yuqing and Ma, Yan and Liu, Pengfei , booktitle =. 2024 , url = on Pith
  47. [67] Mitchell, Eric and Rafailov, Rafael and Sharma, Archit and Finn, Chelsea and Manning, Christopher D. , journal =. 2023 , url = on Pith
  48. [68] Zhang, Yunxiang and Khalifa, Muhammad and Zhang, Lechen and Liu, Xin and Lee, Ayoung and Zhang, Xinliang Frederick and Fatahi Bayat, Farima and Wang, Lu , booktitle =. 2026 , url = on Pith
  49. [69] Cloud, Alex and Le, Minh and Chua, James and Betley, Jan and Sztyber-Betley, Anna and Hilton, Jacob and Marks, Samuel and Evans, Owain , journal =. 2025 , url = on Pith
  50. [70] Khalil, Ali and Kassem, Aly M. and Abdelrazek, Mohamed and Rana, Santu and Rostamzadeh, Negar and Farnadi, Golnoosh , journal =. 2026 , url = on Pith
  51. [71] Andriushchenko, Maksym and Croce, Francesco and Flammarion, Nicolas , booktitle =. 2025 , url = on Pith
  52. [72] Wang, Andy and Mahajan, Parv and Africa, David Demitri and Souly, Alexandra and Taylor, Jordan and Kirk, Robert , journal =. 2026 , url = on Pith
  53. [73] Ge, Haosen and Li, Shuo and Huang, Lianghuan , journal =. 2025 , url = on Pith
  54. [74] Roytburg, Dani and Sridhar, Shreya and Ippolito, Daphne , booktitle =. 2026 , url = on Pith
  55. [75] Zhou, Chunting and Liu, Pengfei and Xu, Puxin and Iyer, Srini and Sun, Jiao and Mao, Yuning and Ma, Xuezhe and Efrat, Avia and others , booktitle =. 2023 , url = on Pith
  56. [76] 2026 , month = feb, howpublished = on Pith
  57. [77] 2026 , month = apr, howpublished = on Pith
  58. [78] Qiu, Zihan and Wang, Zekun and Li, Xiao and Li, Yanpeng and Xu, Yang and Wang, Yixuan and Zhang, Huaqing and Men, Rui and Mao, Bochao and Zhang, Chengruidong and others , journal =. On the Design of. 2026 , url = on Pith
  59. [79] 2026 , month = aug, howpublished = on Pith
  60. [80] arXiv preprint arXiv:2606.19348 , year = on PitharXiv
  61. [81] arXiv preprint arXiv:2609.19969 , year = on PitharXiv
  62. [82] 2026 , month = jul, howpublished = on Pith
  63. [83] Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =. 2023 , url = on Pith
  64. [84] Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph E. and Zhang, Hao and Stoica, Ion , booktitle =. Efficient Memory Management for Large Language Model Serving with. 2023 , doi = on Pith
  65. [85] Zhao, Yanli and Gu, Andrew and Varma, Rohan and Luo, Liang and Huang, Chien-Chin and Xu, Min and Wright, Less and Shojanazeri, Hamid and Ott, Myle and Shleifer, Sam and others , journal =. 2023 , doi = on Pith
  66. [86] Monroe, Burt L. and Colaresi, Michael P. and Quinn, Kevin M. , journal = on Pith
  67. [87] Covington, Michael A. and McFall, Joe D. , journal = on Pith
Cited by 1 Pith paper

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment