HARGO: Heterogeneity-Aware Reward-Guided Optimization for RL Post-Training of LLMs on HPC Tasks

Tiangang Li; Xiangbo Tian

REVIEW 4 major objections 6 minor 16 references

HARGO: Heterogeneity-Aware Reward-Guided Optimization for RL Post-Training of LLMs on HPC Tasks

T0 review · 4 major / 6 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read HARGO reweights each RL response by confidence-modulated advantage so one training loop can handle heterogeneous HPC tasks without task labels.

desk verdict Solid small-scale GRPO reweighting for mixed HPC tasks; primary-metric lead is real in their setup but confounded by unmatched β and single-seed runs. read the letter →

arxiv 2607.28301 v1 pith:ZVXQYOCG submitted 2026-07-30 cs.LG

Tiangang Li , Xiangbo Tian This is my paper

classification cs.LG

keywords reinforcementlearningpost-traininghigh-performancecomputingtaskheterogeneityrewardweightingdomain-specificLLMGRPOdataracedetectionconfidence-modulatedadvantage

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Supervised fine-tuning can teach a language model HPC facts, but the same model that nails data-race yes/no labels still produces long, imprecise answers on factual benchmarks. Reinforcement learning can close that behavior gap by scoring whole answers, yet HPC tasks differ by nearly sixty times in answer length, use three different reward shapes, and start from very different accuracy levels, so methods that treat every sample equally waste gradient on easy batches and under-learn hard ones. HARGO keeps the familiar group-relative RL setup and adds two internal signals: how much a response stands out from its group on reward, and how confident the frozen reference model is about that response. It multiplies the advantage by a bounded confidence factor, then weights the loss by the absolute modulated advantage, with equal weights only when the group is uninformative. On four HPC tasks and nine methods from the same small checkpoint, the paper reports that this scheme leads all three primary metrics it defines, and ablation shows both signals help. A sympathetic reader cares because the method claims better alignment quality from the same compute budget without needing task-type labels or extra models.

What carries the argument

Confidence-modulated advantage: take the usual group-relative advantage A, scale it by (1 + α·c) where c is a sigmoid of the reference model’s mean log-probability relative to an EMA baseline, then set each response’s loss weight proportional to |A_mod| (with equal-weight fallback when advantages vanish). This amplifies informative rollouts while keeping the gradient direction and bounding the boost to at most 30%.

What would settle it

Retrain HARGO and the main online baselines from the same SFT checkpoint with matched KL coefficients, several random seeds, and report means and intervals on WinRate, Data Race F1, and PLP Similarity; if HARGO no longer leads those three primary metrics under matched settings, the central ranking claim fails.

Watch

Extended reading notes

Core claim

The paper claims that HPC task heterogeneity makes uniform-weight online RL such as GRPO suboptimal, and that per-response importance weighting via confidence-modulated advantage—A_mod = A × (1 + α·c) with weights proportional to |A_mod|, α = 0.3, no task labels—yields the best overall alignment quality among nine compared methods on four HPC tasks, leading WinRate (54.62%), Data Race F1 (91.30%), and PLP Similarity (0.8558), with both the discrimination and confidence signals contributing.

Load-bearing premise

That topping three author-chosen primary metrics on one 0.5B checkpoint, one fixed reward, small groups, and one auto-preference pipeline—without multi-seed error bars and with a lower KL coefficient than the closest baseline—is enough to call the method best overall alignment quality for heterogeneous HPC RL.

Editorial extensions

If this is right

Uniform 1/G weighting in group RL is a real failure mode when tasks differ sharply in length, reward shape, and baseline accuracy.
Discrimination from group reward contrast and confidence from reference log-probs can be combined without task-type labels or extra learned models.
Advantage modulation can reallocate a fixed rollout budget toward higher-value responses and improve primary alignment metrics on mixed HPC workloads.
Offline preference methods that rely on auto-generated pairs can lag online weighted methods on these HPC tasks under the paper’s protocol.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same two-signal weighting may transfer to other mixed-format domains (code plus short fact QA plus long explanation) where groups mix easy and hard items.
If the method’s gains come mainly from reallocating gradient rather than extra samples, compute-efficiency curves at fixed wall-clock would be a natural next measurement.
Matching regularization strength across methods before ranking would tighten the comparison the paper already frames as objective-only.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The useful bit here is a clean, label-free fix for GRPO on heterogeneous batches: modulate group advantage by reference-model confidence (A_mod = A × (1 + α c), α=0.3), then weight by |A_mod|. They quantify the heterogeneity (58 imes length, three reward shapes, SFT accuracy 51–100%), run nine methods from one Qwen2.5-0.5B SFT checkpoint with a shared external reward, and ablate d vs c. That is more transparent than most RLHF-variant preprints.

What they do well: evaluation is not circular. Race F1 uses independent yes/no extraction, PLP uses sentence embeddings, WinRate is vs SFT. They openly lose EM/AvgScore to KTO and discuss the mlperf trade-off instead of burying it. Ablation (B1/B2/B3) shows both signals move WinRate and that full modulation beats d alone on the primaries. Math is simple and the algorithm is fully specified; HF data and a checkpoint are linked.

Soft spots are real but proportionate. β is unmatched (HARGO 0.02 vs GRPO 0.04); they even credit the lower KL to “built-in regularization,” so the +1.45 WR / +0.51 F1 edges over GRPO/DrGRPO mix the weighting idea with a looser budget. Everything is single-seed, G=4, 0.5B—margins sit inside plausible noise. Auto-generated preferences for the offline baselines are a weaker control. None of this breaks the internal story; it just means “best overall alignment” is scoped to their metric choice and config, not settled.

This is for people doing domain RL post-training on mixed scientific/HPC workloads who already run GRPO-style loops. Not a theory paper. I would send it to referees: the experiment is controlled enough and the idea is concrete enough to deserve a proper look, with the obvious asks for matched β, multi-seed error bars, and a larger backbone. Worth engaging if you care about small-model HPC assistants; skip if you only track frontier-scale alignment.

Referee Report

4 major / 6 minor

Summary. The paper argues that HPC tasks (binary data-race detection, factual MLPerf QA, and descriptive PLP QA) are too heterogeneous for uniform-weight online RL such as GRPO, and proposes HARGO: per-response importance weights derived from a confidence-modulated group advantage A_mod,i = A_i × (1 + α·c_i), with w_i ∝ |A_mod,i| and equal-weight fallback when advantages vanish. Discrimination d comes from group reward contrast; confidence c from sigmoid-normalized reference-model log-probabilities under an EMA baseline. No task-type labels are required. Starting from a shared Qwen2.5-0.5B SFT checkpoint and a shared task-adaptive reward, the authors compare nine methods and report that HARGO leads three author-defined primary metrics (WinRate 54.62%, Data Race F1 91.30%, PLP Similarity 0.8558), with an ablation (d-only, c-only, full) offered as evidence that the two signals are complementary.

Significance. Task heterogeneity in domain RL post-training is a genuine and under-addressed problem; framing it with measurable dimensions (answer length, reward shape, SFT accuracy) and proposing a label-free, GRPO-compatible reweighting scheme is a useful contribution if the gains are real. Strengths include a shared SFT init and shared external reward across RL methods, evaluation metrics that are not identical to the training reward (independent yes/no extraction for race F1; sentence-embedding PLP similarity; WinRate vs SFT), a transparent ablation table, and release of dataset/checkpoint links. The method itself is simple and cheap (EMA scalar + sigmoid). Significance is currently limited by the 0.5B single-seed setup, unmatched regularization across baselines, and modest absolute margins; if those are cleaned up, the work would be a solid empirical systems paper for domain RL on heterogeneous task suites.

major comments (4)

[Table 3; Supp. Table 8; §6] Table 3 and Supplementary Table 8: HARGO is trained with KL coefficient β=0.02 while GRPO/DrGRPO use β=0.04 and DPO/KTO use β=0.1. Section 6 explicitly attributes the lower β to “built-in regularization” from advantage modulation. The primary-metric ranking (Table 4) therefore confounds the proposed weighting scheme with a looser KL budget. A matched-β re-run of HARGO vs GRPO/DrGRPO (and, ideally, a β sweep) is needed before the claim that A_mod weighting establishes best overall alignment quality can be treated as isolated.
[Tables 4–7; §5 Main Results] All headline numbers in Tables 4–7 are single-run with no multi-seed means or error bars. The WinRate edge over GRPO is +1.45 points and F1 over DrGRPO is +0.51; at G=4 on a 0.5B model these margins are within plausible seed noise. The central ranking claim (“best on all three primary metrics” / “best overall alignment quality”) is not statistically supported as written. Report at least 3 seeds with mean±std for the primary metrics, or another clear uncertainty estimate.
[Table 7; §5 Ablation Study] Table 7 ablation: B1 (w ∝ d only) already reaches F1 91.37 and WinRate 53.34, i.e., most of the F1 lead and a large fraction of the WinRate lead over GRPO. Full HARGO (B3) improves WinRate and PLP but slightly lowers EM/AvgScore and F1 vs B1. The prose claim that both signals “independently contribute” and that the combination is necessary for the ranking should be qualified: under matched compute, d alone appears to carry detection performance, and the complementarity argument rests mainly on WinRate/PLP. Clarify what is gained uniquely by modulation versus discrimination-only weighting, especially once β is matched.
[§5 Evaluation metrics; Table 4; §6] Sections 5–6 elevate three author-chosen primary metrics and discount EM/AvgScore (where KTO leads) as “auxiliary” measuring “reproduction fidelity rather than alignment quality.” That hierarchy is defensible only if justified a priori and applied consistently; as written it risks post-hoc ranking. Either pre-register the primary set and report a single aggregate (e.g., average rank across all five metrics), or show that conclusions are stable under reasonable alternative weightings of the metric suite—including mlperf accuracy where HARGO is third.

minor comments (6)

[Abstract; §1] Widespread missing spaces after periods/commas and concatenated words throughout (e.g., Abstract, §1, Figure 1 caption). A full copy-edit pass is needed before camera-ready.
[Supp. B Algorithm 1; Eqs. (2)–(5)] Algorithm 1 (Supp. B) applies the policy gradient to A_mod,i inside L_i while the main text (Eqs. 2–5) defines the clipped surrogate on A_i and uses A_mod only for weights. State clearly whether the surrogate advantage is A or A_mod; the two versions are not equivalent.
[Figure 1] Figure 1 panels are informative but axis labels and density curves are hard to read in the compiled text; ensure vector figures and consistent task colors across (a)–(c).
[Table 6; Table 5] Table 6 DPO row shows TP=104, FN=52 and a large Fortran collapse; a one-sentence note on preference-data quality for Fortran would help readers interpret the offline baselines.
[§2] Related Work cites many 2025–2026 HPC-LLM papers; verify final bibliographic details and page numbers for camera-ready consistency with arXiv versions.
[§4 Per-Response Weights] Hyperparameter α=0.3 is said to be chosen by a sweep over {0.1, 0.3, 1.0} but sweep results are not shown; a small table or sentence with the three WinRate/F1 values would strengthen the design-decision paragraph.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HARGO is an empirical weighting design evaluated on independently computed metrics, not a self-defining prediction chain.

full rationale

The paper proposes a concrete training objective change on top of GRPO—modulated advantage A_mod,i = A_i × (1 + α·c_i), discrimination d_i ∝ |A_mod,i|, weights w_i ∝ d_i—and evaluates it empirically against eight baselines on four HPC tasks. The load-bearing claims are comparative leaderboard results (WinRate, Data Race F1, PLP Similarity), not first-principles predictions. Training and evaluation are not the same object by construction: the shared reward R is an external fixed scorer used identically by all RL methods; Data Race F1 is computed via independent word-boundary yes/no extraction; PLP Similarity uses an external sentence-transformer cosine; WinRate is preference vs the SFT baseline. α and β are hyperparameters chosen by sweep/design, not fitted quantities renamed as predictions. There is no self-citation uniqueness theorem, no ansatz smuggled from prior author work as external fact, and no definitional loop equating inputs to claimed outputs. Concerns about unmatched β, single-seed runs, or primary-vs-auxiliary metric emphasis are experimental-validity issues, not circularity under the stated criteria. The derivation chain is self-contained empirical method + benchmark comparison.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central ranking rests on standard online RL math (group-relative advantage, clipped surrogate, KL to reference), an author-specified task-adaptive reward, hand-chosen modulation strength α and related training knobs, and the premise that three primary metrics define ‘best overall alignment’ on this heterogeneous suite. No new physical entities; free parameters are optimization/design choices fitted or swept on this run.

free parameters (5)

confidence coefficient α = 0.3
Scales how much reference confidence boosts advantage magnitude; swept over {0.1, 0.3, 1.0} and fixed at 0.3 for main results.
KL coefficient β (HARGO) = 0.02
Set lower than GRPO (0.04) and preference methods (0.1); authors argue modulation supplies regularization. Directly affects stability/quality trade-off in reported runs.
EMA decay ρ for ref_logp_global = 0.9
Controls moving baseline for confidence signal c; fixed without extensive sensitivity in main text.
group size G, temperature T, epochs, lr, max new tokens = G=4, T=0.6, 3 epochs, lr=1e-5, max_new_tokens=64
Standard RLHF knobs that shape reward variance and learning signal; G=4 and T=0.6 are small and material at 0.5B.
task-adaptive reward tiers and penalties = as specified in §5 Reward function
mlperf partial-match tiers (1.0/0.5/0.2/0.1), plp embedding similarity, +0.05 format bonus, -0.001/char length penalty beyond 200 define the optimization target shared by all RL methods.

assumptions (5)

domain assumption Group-relative normalized advantages as in GRPO are a valid learning signal without a learned value function at 0.5B scale.
Imported from Shao et al. 2024 and used as the base A_i before modulation (§4 Preliminaries).
ad hoc to paper Frozen SFT reference log-probabilities are a meaningful confidence proxy for whether an error contradicts learned HPC knowledge.
Motivates c_i = σ(ref_logp_i - ref_logp_global); not independently validated outside this weighting scheme (§4 Confidence Signal).
domain assumption A single fixed heuristic reward R (exact match / keyword tiers / MiniLM cosine) adequately represents task-appropriate HPC behavior for ranking methods.
All RL methods optimize the same R; primary PLP metric also uses MiniLM, creating partial dependence on embedding geometry (§5).
domain assumption Automatically labeled SFT-vs-SFT preference pairs are a fair substrate for DPO/KTO/SimPO baselines.
Preference data generation protocol in §5; weak pairs may systematically disadvantage offline methods in the nine-way comparison.
standard math Standard clipped policy-gradient + KL math and sigmoid/EMA normalizations behave as usual.
Equations (1)–(11) and Algorithm 1.

invented entities (1)

HARGO confidence-modulated advantage A_mod and discrimination weight w ∝ |A_mod|
purpose: Reallocate per-response gradient mass under task heterogeneity without task labels.
Defined in §4 as the paper’s methodological object; evaluated only inside this experimental loop.

how reviews work

0 comments

Cite this review

Pith. "Pith review of HARGO: Heterogeneity-Aware Reward-Guided Optimization for RL Post-Training of LLMs on HPC Tasks." pith.science (2026). https://pith.science/paper/ZVXQYOCG

@misc{pith2026260728301,
  author       = {Pith},
  title        = {Pith review of: HARGO: Heterogeneity-Aware Reward-Guided Optimization for RL Post-Training of LLMs on HPC Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZVXQYOCG}},
  note         = {Machine review of arXiv:2607.28301}
}

read the original abstract

Supervised fine-tuning (SFT) can equip large language models (LLMs) with domain knowledge for high-performance computing (HPC) tasks such as data race detection and benchmark question answering. However, knowledge alone does not guarantee task-appropriate behavior: the same SFT model that correctly classifies 88.65\% of C/C++ data race samples produces verbose, imprecise answers to factual queries, with 65.9\% of MLPerf responses exceeding 40 characters. Reinforcement learning (RL) post-training addresses this gap by optimizing for task-specific rewards rather than token-level imitation. Yet HPC tasks exhibit extreme heterogeneity, with binary classification, factual QA, and semantic generation differing by 58x in answer length, spanning three distinct reward distributions, and showing widely varying SFT accuracy. This makes uniform-weight RL methods such as GRPO suboptimal. We propose HARGO, Heterogeneity-Aware Reward-Guided Optimization, which introduces per-response importance weighting via confidence-modulated advantage: computing a discrimination signal from group-level reward contrast and a confidence signal from reference model log-probabilities, then modulating the advantage before computing per-response weights, without requiring task-type labels. Across four HPC tasks and nine methods, HARGO achieves the best performance on all three primary metrics: WinRate 54.62\%, Data Race F1 91.30\%, and PLP Similarity 0.8558. Ablation confirms complementary contributions from both signals. HARGO establishes the best overall alignment quality among compared methods for heterogeneous HPC tasks.

Figures

Figures reproduced from arXiv: 2607.28301 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 6 linked inside Pith

[2]

InIWOMP 2023, 18–33

LM4HPC: Towards Effective Lan- guage Model Application in High-Performance Computing. InIWOMP 2023, 18–33. Kadosh,T.;etal.2024.MonoCoder:Domain-SpecificCode Language Model for HPC Codes and Tasks. InIEEE HPEC 2024, 1–7. Chen, L.; Bhattacharjee, A.; Ahmed, N. K.; Hasabnis, N.; Oren,G.;Vo,V.;andJannesari,A.2024.OMPGPT:AGen- erativePre-trainedTransformerMode...

2023
[4]

ChatHPC: Building the Foundations for a Productive and TrustworthyAI-AssistedHPCEcosystem.InSC 2025,458–

2025
[6]

InICPP-W 2025, 1–8

Comprehensive Evaluation of LLMs in HPC Code Perfor- mance Optimization. InICPP-W 2025, 1–8. Nader, N.; Diehl, P.; Brandt, S.; and Kaiser, H

2025
[9]

InSC-W 2023, 215–223

Data Race Detection Using Large Language Models. InSC-W 2023, 215–223. Ljaljevic,S.;Jorba,J.;andIserte,S.2026.ExploringtheRole ofLargeLanguageModelsinHigh-PerformanceComputing Programming: A Survey.Future Generation Computer Sys- tems, 184:108618. Chen,L.;Ahmed,N.K.;Dutta,A.;Bhattacharjee,A.;Yu,S.; Mahmud, Q. I.; et al

2023
[10]

Zhang, S.; Zhao, J.; Yu, Q.; Xia, C.; Wang, Z.; Feng, X.; and Cui, H

The Landscape and Challenges of HPC Research and LLMs.Preprint, arXiv:2402.02018. Zhang, S.; Zhao, J.; Yu, Q.; Xia, C.; Wang, Z.; Feng, X.; and Cui, H

arXiv
[11]

Kadosh,T.;Hasabnis,N.;Vo,V.A.;Schneider,N.;Krien,N.; Wasay,A.;etal.2023.ScopeIsAllYouNeed:Transforming LLMs for HPC Code.Preprint, arXiv:2308.09440

OMPar: Automatic Parallelization with AI-Driven Source-to-Source Compila- tion.Preprint, arXiv:2409.14771. Kadosh,T.;Hasabnis,N.;Vo,V.A.;Schneider,N.;Krien,N.; Wasay,A.;etal.2023.ScopeIsAllYouNeed:Transforming LLMs for HPC Code.Preprint, arXiv:2308.09440. Chaturvedi,A.;Nichols,D.;Singh,S.;andBhatele,A.2025. HPC-Coder-v2:StudyingCodeLLMsAcrossLow-Resource ...

arXiv 2023
[12]

Meng,Y.;Xia,M.;andChen,D.2024.SimPO:SimplePref- erenceOptimizationwithaReference-FreeReward.InICML

2024
[13]

DeepSeek-AI.2025.DeepSeek-R1:IncentivizingReasoning Capability in LLMs via Reinforcement Learning.Nature, 645:633–638

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.Preprint, arXiv:2402.03300. DeepSeek-AI.2025.DeepSeek-R1:IncentivizingReasoning Capability in LLMs via Reinforcement Learning.Nature, 645:633–638. Liu,Z.;Chen,C.;Li,W.;Qi,P.;Pang,T.;Du,C.;Lee,W.S.; and Lin, M

arXiv 2025

Show all 16 references

[14]

Touvron, H.; Martin, L.; Stone, K.; Albert, P.; Almahairi, A.;Babaei, Y.;etal.2023.LLaMA2:OpenFoundationand Fine-Tuned Chat Models.Preprint, arXiv:2307.09288

GPT-4 Technical Report.Preprint, arXiv:2303.08774. Touvron, H.; Martin, L.; Stone, K.; Albert, P.; Almahairi, A.;Babaei, Y.;etal.2023.LLaMA2:OpenFoundationand Fine-Tuned Chat Models.Preprint, arXiv:2307.09288. Qwen Team

2023 arXiv
[15]

Liao, C.; Lin, P.-H.; Asplund, J.; Schordan, M.; and Karlin, I.2017.DataRaceBench:ABenchmarkSuiteforSystematic Evaluation of Data Race Detection Tools

Qwen2.5 Technical Report.Preprint, arXiv:2412.15115. Liao, C.; Lin, P.-H.; Asplund, J.; Schordan, M.; and Karlin, I.2017.DataRaceBench:ABenchmarkSuiteforSystematic Evaluation of Data Race Detection Tools. InSC

2017 arXiv
[81]

Preprint, arXiv:2501.14733

Miyashita,Y.;Tung,P.K.M.;andBarthelemy,J.2025.LLM asHPCExpert:ExtendingRAGArchitectureforHPCData. Preprint, arXiv:2501.14733. Chen, L.; Ding, X.; Emani, M.; Vanderbruggen, T.; Lin, P.- H.; and Liao, C

2025 arXiv
[2019]

2020 (ongoing)

MLCommons. 2020 (ongoing). MLPerf Training Bench- mark. https://mlcommons.org/benchmarks/training/. Supplementary Document This supplementary document provides: (1) the complete training configurations for all baseline methods evaluated in the main paper, and (2) the pseudocod...

2020
[2023]

InSC-W 2023, 951–960

HPC-GPT: Inte- grating Large Language Model for High-Performance Com- puting. InSC-W 2023, 951–960. Chen, L.; et al

2023
[2024]

InSC-W 2024, 19–24

ChatBLAS: The First AI-Generated and Portable BLAS Library. InSC-W 2024, 19–24. Godoy, W. F.; Valero-Lara, P.; Teranishi, K.; Balaprakash, P.;andVetter,J.S.2024.LargeLanguageModelEvaluation for High-Performance Computing Software Development. Concurrency and Computation: Pract...

2024
[2025]

InSC-W 2025, 727–739

AskHPC: A ChatBot for High Performance Computing User Support. InSC-W 2025, 727–739. Valero-Lara, P.; Young, A.; Vetter, J. S.; et al

2025
[2026]

InHigh Performance Com- puting (ISC 2025), Springer, 626–638

LLM & HPC: Benchmarking DeepSeek’s Performance in High- Performance Computing Tasks. InHigh Performance Com- puting (ISC 2025), Springer, 626–638. Alsofyani, M.; and Wang, L

2025

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.

[1] [2]

InIWOMP 2023, 18–33

LM4HPC: Towards Effective Lan- guage Model Application in High-Performance Computing. InIWOMP 2023, 18–33. Kadosh,T.;etal.2024.MonoCoder:Domain-SpecificCode Language Model for HPC Codes and Tasks. InIEEE HPEC 2024, 1–7. Chen, L.; Bhattacharjee, A.; Ahmed, N. K.; Hasabnis, N.; Oren,G.;Vo,V.;andJannesari,A.2024.OMPGPT:AGen- erativePre-trainedTransformerMode...

2023

[2] [4]

ChatHPC: Building the Foundations for a Productive and TrustworthyAI-AssistedHPCEcosystem.InSC 2025,458–

2025

[3] [6]

InICPP-W 2025, 1–8

Comprehensive Evaluation of LLMs in HPC Code Perfor- mance Optimization. InICPP-W 2025, 1–8. Nader, N.; Diehl, P.; Brandt, S.; and Kaiser, H

2025

[4] [9]

InSC-W 2023, 215–223

Data Race Detection Using Large Language Models. InSC-W 2023, 215–223. Ljaljevic,S.;Jorba,J.;andIserte,S.2026.ExploringtheRole ofLargeLanguageModelsinHigh-PerformanceComputing Programming: A Survey.Future Generation Computer Sys- tems, 184:108618. Chen,L.;Ahmed,N.K.;Dutta,A.;Bhattacharjee,A.;Yu,S.; Mahmud, Q. I.; et al

2023

[5] [10]

Zhang, S.; Zhao, J.; Yu, Q.; Xia, C.; Wang, Z.; Feng, X.; and Cui, H

The Landscape and Challenges of HPC Research and LLMs.Preprint, arXiv:2402.02018. Zhang, S.; Zhao, J.; Yu, Q.; Xia, C.; Wang, Z.; Feng, X.; and Cui, H

arXiv

[6] [11]

Kadosh,T.;Hasabnis,N.;Vo,V.A.;Schneider,N.;Krien,N.; Wasay,A.;etal.2023.ScopeIsAllYouNeed:Transforming LLMs for HPC Code.Preprint, arXiv:2308.09440

OMPar: Automatic Parallelization with AI-Driven Source-to-Source Compila- tion.Preprint, arXiv:2409.14771. Kadosh,T.;Hasabnis,N.;Vo,V.A.;Schneider,N.;Krien,N.; Wasay,A.;etal.2023.ScopeIsAllYouNeed:Transforming LLMs for HPC Code.Preprint, arXiv:2308.09440. Chaturvedi,A.;Nichols,D.;Singh,S.;andBhatele,A.2025. HPC-Coder-v2:StudyingCodeLLMsAcrossLow-Resource ...

arXiv 2023

[7] [12]

Meng,Y.;Xia,M.;andChen,D.2024.SimPO:SimplePref- erenceOptimizationwithaReference-FreeReward.InICML

2024

[8] [13]

DeepSeek-AI.2025.DeepSeek-R1:IncentivizingReasoning Capability in LLMs via Reinforcement Learning.Nature, 645:633–638

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.Preprint, arXiv:2402.03300. DeepSeek-AI.2025.DeepSeek-R1:IncentivizingReasoning Capability in LLMs via Reinforcement Learning.Nature, 645:633–638. Liu,Z.;Chen,C.;Li,W.;Qi,P.;Pang,T.;Du,C.;Lee,W.S.; and Lin, M

arXiv 2025