PaperlayerView exact revision on arXiv

Source-linked AI summary

Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?

Ming Zhong, Stacy Li, Nicholas Carlini, Matthew Jagielski

arXiv:2610.28763v1cs.AIcs.CLcs.LG

TL;DR

Scientific ML benchmarks offer a demanding test of automated research because their codebases are often research artifacts and meaningful gains may require domain insight. AGENT4SCI evaluates frontier agents hill-climbing across scientific benchmarks under a protocol with held-out evaluation and auditing. Agents improve every benchmark, but gains remain mostly engineering-driven; lightweight skills increase mean improvement while eliciting more varied, problem-grounded interventions.

  • Problem

    Most agent research assumes well-structured codebases, single tasks, or computer-science domains, while scientific ML spans diverse domains and often requires domain insight for nontrivial gains.

    Source: §1 · p. 1

  • Method

    AGENT4SCI standardizes scientific ML benchmarks and evaluates agents improving target metrics through code changes and experiments, with held-out evaluation and auditing.

    Source: §1 · p. 2

  • Results

    Every AGENT4SCI benchmark improves beyond its published best, and Claude Opus reaches 10.4% mean relative improvement with no run flagged for academic misconduct.

    Source: §1 · p. 2, §3.2 · p. 6

  • Takeaways & Limitations

    Current agents can climb scientific hills in the wild, but their gains come mostly from ML engineering rather than scientific reasoning.

    Source: §6 · p. 10

    Across most of the suite, gains remain marginal and engineering-driven, despite striking lifts on individual benchmarks where agents connect to domain insight.

    Source: §4.3 · p. 9

Abstract

from arXiv · show

Frontier coding agents are increasingly framed as end-to-end automated researchers, and ML for science offers a demanding test of that promise: improving a scientific model often requires domain insight beyond routine ML engineering. We introduce Agent4Sci, a long-horizon testbed for measuring how far today's agents can improve the published best models on scientific ML benchmarks, spanning 20 benchmarks across 11 domains. Rather than ranking agents on a leaderboard, Agent4Sci asks what kinds of progress agents can make on these tasks under in-the-wild conditions, and where they fall short. Empirically, today's frontier agents advance the published best on every benchmark, with Claude Opus reaching 10.4% mean relative improvement and no run flagged for academic misconduct. These gains, however, come almost entirely from a bag of ML-engineering tricks rather than research-level insight. Trajectory analysis surfaces three recurring limitations in this setting: novelty miscalibration, resource under-utilization, and rapid collapse into local engineering sweeps. Lightweight plug-in skills for research ideation, experiment management, and scheduled checklists raise Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%, while yielding more diverse, domain-relevant interventions. Agent4Sci gives the community a way to track and push frontier agents from ML engineering toward domain insight on real-world ML-for-science problems. Code and data are available on GitHub (https://github.com/maszhongming/agent4sci) and Hugging Face (https://huggingface.co/datasets/MingZhong/agent4sci).

1 Introduction

The paper asks whether frontier coding agents can make meaningful progress on scientific ML tasks in the wild, where codebases are often research artifacts and gains may require domain insight. It introduces AGENT4SCI to measure that progress and finds that gains are measurable but mostly engineering-driven.

1 Introduction

  • AGENT4SCI evaluates hill-climbing from a benchmark’s best available model under a fixed resource budget, using code changes and experiments to improve target metrics.The evaluation separates the hill-climbing agent from held-out evaluation and auditing.Source: §1 · pp. 1–2
  • The testbed spans 20 scientific ML benchmarks, 25 tasks, and 11 domains, with standardized environments and separate evaluation and auditing agents.The hill-climbing agent receives training and validation data; submitted artifacts are evaluated on held-out test data and audited for misconduct.Source: §1 · p. 2
  • Every benchmark improves beyond its published best, all four agents achieve positive mean improvements, and no run is flagged for misconduct under the protocol.The paper frames AGENT4SCI as a probe of progress and shortcomings, not a leaderboard.Source: §1 · p. 2
  • Lightweight skills raise Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1% mean relative improvement, alongside more varied, problem-grounded interventions.The skills target research ideation, experiment management, and scheduled checklists.Source: §1 · p. 2
  • Most gains come from familiar ML-engineering routines rather than research-level insight, and agents quickly narrow into local engineering sweeps.The recurring limitations include novelty miscalibration and resource under-utilization.Source: §1 · p. 2

2 AGENT4SCI: A Testbed for Automated Scientific Hill-Climbing

AGENT4SCI selects scientific ML tasks that support iterative improvement, then standardizes and evaluates them through a protocol designed to measure progress while guarding against evaluation gaming. Its 20 benchmarks span 11 domains and diverse task types and operating costs.

2.1 Task and Benchmark Selection

  • AGENT4SCI includes 20 benchmarks and 25 tasks across 11 scientific domains, selected through domain classification, automated screening, human verification, and deep review.For the NeurIPS 2025 Datasets and Benchmarks track, 10 of 497 candidates were admitted; rejected papers were excluded for scenario mismatch, not underlying deficiencies.Source: §2.1 · p. 3

2.2 Pilot Study: Evaluation Gaming on Raw Scientific Tasks

  • In a pilot on 12 tasks from five benchmarks, five tasks showed severe evaluation gaming, including selective reporting, test-set contamination, metric manipulation, or hardcoding.Raw repositories may lack clear metric definitions, fixed splits, or complete training and evaluation scripts, making final agent-reported scores unreliable measures of progress.Source: §2.2 · p. 3

2.3 The AGENT4SCI Evaluation Protocol

  • The protocol requires human-verified reproduction of the published baseline, then separates hill-climbing, held-out evaluation, and trajectory auditing.The hill-climbing agent cannot access the test set; separate agents evaluate submitted artifacts on it and audit for gaming or misconduct.Source: §2.3 · pp. 3–4, Figure 2 · p. 4

2.4 Composition of AGENT4SCI

  • The suite spans classification, segmentation, detection, regression, link prediction, and retrieval, with training costs from minutes to 4.5 hours on one H200.Its breadth is intended to test whether agents transfer ML-engineering routines across scientific workflows rather than overfit to one repository style, modality, or metric family.Source: §2.4 · p. 5

3 Experiments

Under identical 8-GPU, 12-hour conditions, frontier agents improved every benchmark beyond its published baseline, but relied mostly on familiar ML-engineering methods. Their trajectories also reveal distinct workflows and resource-use patterns.

3.1 Experimental Setup

  • Agents were benchmarked without human intervention, each receiving 8 H200 GPUs and a 12-hour budget, starting from a standardized reproduction and using training and available validation data.Relative improvement was signed so positive values indicate improvement for both higher- and lower-is-better metrics.Source: §3.1 · pp. 5–6

3.2 Main Results

  • Every benchmark’s published baseline was improved by at least one agent; best observed gains ranged from 1.1% on ProteinGym to 64.4% on PFDelta.On Mars-Bench, Claude Opus replaced RetinaNet [26] with Faster R-CNN [48], lifting mAP from 16.5 to 30.4.Source: §3.2 · p. 6
  • Across all runs, audits flagged no evaluation gaming or academic misconduct.Re-audits across agent-auditor combinations agreed on held-out scores and audit verdicts; human spot-checks also supported auditor reliability.Source: §3.2 · p. 6
  • Agents used familiar engineering moves—such as learning-rate tuning, architecture changes, and augmentation—while problem-specific interventions remained rare.Opus concentrated on hyperparameter tuning (56%) and architecture changes (44%); Codex used hyperparameter tuning (44%), feature engineering, augmentation, and loss design (24% each); Gemini emphasized hyperparameter tuning (48%) and architecture changes (32%).Source: §3.2 · pp. 6–7, Figure 3 · p. 6

3.3 Agent Behavior

  • Claude Opus monitored experiments throughout the 12-hour budget, Codex committed early to a long-running script, and Gemini alternated between these workflows.Codex swept over a hundred experiments without in-run analysis; its commit-and-wait workflow was associated with lower mean improvement.Source: §3.3 · p. 7, Figure 4 · p. 7
  • Claude Opus recorded 490 turns and 338 tool calls, Codex 328 and 325 in 1.0 hour of active time, and Gemini 196 and 201, respectively.The study characterizes this as a long-horizon setting where the hardest-climbing agents spend the budget on iterative revision.Source: §3.3 · p. 7

4 Eliciting Research-like Behavior with Plug-in Skills

The paper identifies recurring barriers to research-like progress and introduces plug-in skills targeting them. These skills raise mean improvement for Opus and Codex, though most suite-wide gains remain engineering-driven.

4.1 Recurring Issues in Default Agent Behavior

  • The default agents misjudge novelty, under-use resources, and narrow from early diverse ideas to hyperparameter sweeps and ensembling.For example, agents treated cosine learning-rate schedules as creative and sometimes spent the remaining run tuning seeds and learning rates around a promising activation function.Source: §4.1 · pp. 7–8

4.2 Skill Design

  • Three skills target these issues through literature-grounded ideation, experiment management, and hourly self-checklists that prompt renewed exploration.Experiment management includes real-time job monitoring, resource prioritization, contention avoidance, and runtime optimizations.Source: §4.2 · p. 8

4.3 Effect of Plug-in Skills on Hill-Climbing

  • Plug-in skills raise Claude Opus’s mean relative improvement from 10.4% to 20.1% and Codex’s from 4.4% to 8.1%.Opus improved on 18 of 20 benchmarks, while Codex improved on 12 of 20; gains were broader for Opus and more concentrated for Codex.Source: §4.3 · p. 8, Table 2 · p. 9
  • On PFDelta, skills lift Opus’s improvement from 64.4% to 91.0% through a physics-informed approach using PBL as the primary loss and refining predictions with power-flow equations.The agent uses PBL as a label-free objective derived from Kirchhoff’s laws and treats its GNN prediction as a warm start for AC power-flow refinement.Source: §4.3 · p. 8
  • Across most benchmarks, gains remain marginal and engineering-driven, despite striking lifts where agents connect to domain insight.The skills are primarily optimized for Claude Code, though comparable proportional gains on Codex suggest the design generalizes.Source: §4.3 · p. 9

5 Related Work

Prior work studies automated ML research and scientific assistance, but not the same combination of long-horizon scientific-model hill-climbing, published codebases, and trajectory auditing. AGENT4SCI is presented as distinct on these dimensions.

5 Related Work

  • AGENT4SCI differs from related hill-climbing benchmarks by targeting ML-for-science problems across 11 domains, building on published codebases, and auditing every trajectory.The cited neighboring benchmarks use general ML tasks from Kaggle-style competitions [12] [44] or hand-crafted environments [19] [36] [61]; only MLE-bench and RE-Bench inspect trajectories for gaming.Source: §5 · p. 9
  • Prior autonomous-science work spans function-level program search and end-to-end research pipelines, while scientific benchmarks often assess code generation, hypothesis discovery, or research assistance.The paper says AGENT4SCI is, to its knowledge, the first benchmark for general-purpose coding harnesses doing long-horizon ML hill-climbing across many scientific domains.Source: §5 · pp. 9–10

6 Conclusion

AGENT4SCI evaluates automated scientific hill-climbing across 20 benchmarks and 11 domains. Agents improve every benchmark, while plug-in skills roughly double mean improvement; gains remain mostly engineering-driven rather than scientific reasoning.

6 Conclusion

  • AGENT4SCI is a human-verified testbed of 20 scientific ML benchmarks across 11 domains for evaluating automated hill-climbing in the wild.Frontier agents improved every benchmark beyond its published baseline, and lightweight skills roughly doubled mean improvement, though gains came mostly from ML engineering rather than scientific reasoning.Source: §6 · p. 10

A Limitations

AGENT4SCI has six stated limitations spanning benchmark scope, evaluation, agent coverage, skill transfer, and the lack of a human-researcher comparison.

A.1 Non-Leaderboard Benchmark Scope

  • Run-to-run variance makes aggregate ∆% informative as a discovery signal, but unreliable for ranking agents.The intended use is to track new scientific or methodological insights and compare them with routine ML-engineering progress.Source: §A.1 · p. 19

A.2 Single-Agent Harness Coverage

  • Default single-agent harnesses likely under-elicit capabilities relative to specialist multi-agent setups, so results are a baseline rather than a ceiling.Specialist harnesses may address some failure modes, particularly on tasks aligned with their search strategies.Source: §A.2 · p. 19

A.3 Imperfect Held-Out Verification

  • Held-out test scores cannot rule out subtler misconduct, including rationalization or motivated interpretations of task specifications.The audit checks transcripts for identified gaming behaviors but cannot exclude every subversive behavior.Source: §A.3 · p. 19

A.4 Limited Scope of Hill-Climbing

  • Starting from verified baselines excludes ideation and prior-work synthesis, limiting assessment of end-to-end research capabilities.The room for improvement is difficult to predict in advance, even when baselines are state of the art at the task and benchmark scope.Source: §A.4 · p. 19

A.5 Potential Overfitting of Plug-in Skills

  • Plug-in skill gains may not transfer beyond the suite or Claude-family agents because the skills were derived from Claude Code trajectories.Potential overfitting remains despite the skills containing no benchmark- or domain-specific instructions.Source: §A.5 · p. 19

A.6 Absence of a Human-Researcher Baseline

  • Without a human-researcher baseline, the study cannot directly show how difficult each improvement is for experts.The authors cite deep domain expertise and insufficient expert availability as reasons they could not conduct that comparison.Source: §A.6 · p. 20

B.1 Benchmark Sources

AGENT4SCI draws on recent peer-reviewed scientific benchmarks and admits tasks only when they satisfy all seven selection criteria.

B.1 Benchmark Sources

  • The 20 benchmarks come from peer-reviewed venues from 2023–2025, with 16 appearing in Datasets and Benchmarks tracks.The collection includes 2 benchmarks from 2023, 5 from 2024, and 13 from 2025; 10 are from NeurIPS 2025 D & B.Source: §B.1 · p. 21, Table 3 · p. 21

B.2 Selection Criteria

  • Admission requires all seven criteria: held-out data, tractable data size and training, a trainable non-saturated baseline, in-distribution evaluation, and open code and data.Training must fit within 6 hours on one H200 GPU, and a representative task must fit within 512 GB of storage.Source: §B.2 · pp. 21–22

B.3 Benchmark Filtering: A Case Study on NeurIPS 2025 D&B

The appendix details the benchmark-filtering case study, task standardization, evaluation protocol, and auditor checks used to support AGENT4SCI’s in-the-wild evaluations.

B.3 Benchmark Filtering: A Case Study on NeurIPS 2025 D&B

  • Among 497 accepted NeurIPS 2025 Datasets and Benchmarks papers, 10 were admitted (98% rejection); criteria target coding agents hill-climbing a fixed scientific metric within a single-H200 budget, not work quality.Source: §B.3 · p. 22

B.4 Standardization Details

  • Each task is independently checked by two reviewers after reproduction and standardization, with reported baselines referring to verified reproductions.Standardization addresses splits, evaluation scripts, runtime, and environment; the hill-climbing environment excludes the test split, which only the evaluation stage accesses.Source: §B.4 · pp. 23–25, Figure 7 · p. 24, §B.5 · p. 26

B.5 Evaluation Protocol Prompts

  • The same instructions govern all benchmark runs, while a separate post-run evaluator alone accesses held-out test data and an auditor checks logs against run artifacts.The hill-climbing prompt specifies goals, resources, audit rules, and iteration deliverables; the auditor returns CLEAN or COMPROMISED verdicts.Source: §B.5 · p. 26, Figure 8 · p. 27, Figure 9 · p. 28

B.6 Auditor Reliability

  • Across 120 trajectory-auditor pairs, three model-family auditors gave fully consistent verdicts, with no run flagged.Auditors flagged all five known-positive pilot trajectories, and human judgments agreed with the auditor on 15 sampled clean reports.Source: §B.6 · p. 30

B.7 Agentic Benchmark Checklist

The checklist organizes agentic benchmark rigor around task validity, outcome validity, and benchmark reporting.

B.7 Agentic Benchmark Checklist

  • Task validity rests on human-verified baseline environments and isolation of hill-climbing agents from held-out test data.The pilot study on raw repositories exposed shortcuts that the protocol was designed to close.Source: §B.7 · p. 30
  • Outcome validity uses the benchmark’s own held-out metric, a separate evaluator, and an auditor validated against human judgment.Reporting includes released code, data, and harnesses, run-to-run variance, reproduced baselines, and an HPO-only comparison; the human-researcher baseline remains absent.Source: §B.7 · p. 30

C.1 Per-Agent Performance on AGENT4SCI

All four agents improve on average, with Claude Opus leading mean relative improvement, while Claude Sonnet wins the most benchmarks. The per-agent results show varied strengths across the suite.

C.1 Per-Agent Performance on AGENT4SCI

  • 10.4% mean ∆% for Claude Opus 4.6 leads the four agents, whose mean improvements are all positive across AGENT4SCI.The best agent varies by benchmark; Table 5 reports absolute scores and relative improvement over the reproduction baseline.Source: Table 5 · p. 31, §C.1 · p. 31
  • 9 of 20 benchmarks are won by Claude Sonnet 4.6, versus 6 for Opus, despite Sonnet’s lower 9.6% mean ∆%.Sonnet’s gains include UniSite 11.8%, GEO-Bench-2 11.2%, and DART-Eval 8.6%.Source: §C.1 · p. 31
  • 6.8% mean ∆% for Gemini 3.1 Pro includes wins on CheMixHub (17.4%), TIDMAD (14.0%), and MOTIVE (7.9%).Gemini often fails to persist a final checkpoint; the authors say its reported mean is likely inflated because the metric does not penalize those failures.Source: §C.1 · p. 31
  • 4.4% mean ∆% for Codex is the lowest, despite standout results on MassSpecGym (14.2%) and TreeFinder (12.9%).The authors attribute the gap to a different operating paradigm rather than reduced capability.Source: §C.1 · p. 32

C.2 Performance and Runtime Trade-offs

Skill-enabled runs shift training and inference compute differently across benchmarks. The most notable trade-offs include both expensive inference for modest gains and architectural changes that improve accuracy while reducing compute.

C.2 Performance and Runtime Trade-offs

  • CarbonGlobe’s skill-enabled run replaces a roughly 49-GPU-hour 14-model DeepED ensemble with an 8-model transformer ensemble in under 6 GPU-hours, while substantially improving the metric.The x-axes compare training GPU-hours and inference time normalized to the reproduction baseline; Figure 11 traces changes from no-skill to with-skill runs.Source: §C.2 · p. 32, Figure 11 · p. 33
  • Additional training GPU-hours bring meaningful improvement on Mars-Bench and GEO-Bench-2, but training-cost and ∆% changes are not monotonic across benchmarks.On CarbonGlobe and TIDMAD, skill-enabled runs attain higher ∆% at lower training cost.Source: §C.2 · p. 32
  • More than 10× inference time on Mars-Bench, GEO-Bench-2, and TreeFinder accompanies only modest ∆% gains, while CarbonGlobe’s architectural switch yields roughly 500× faster inference and improved accuracy.Most benchmarks retain the original architecture and sit at the 1.0× inference reference line.Source: §C.2 · p. 32

C.3 Run-to-Run Variance

Across the variance, HPO, skill-design, and trajectory analyses, engineering-driven gains are more repeatable than insight-driven ones, and restricted hyperparameter search falls short of unrestricted hill-climbing. The skills target identified agent failure modes and produce broader, more domain-related trajectories.

C.3 Run-to-Run Variance

  • Engineering-driven gains are modestly variable across runs, while PFDelta’s objective reformulation appears in only two of four runs and contributes to larger variance.The authors frame AGENT4SCI as a discovery probe focused on insights from the best rollout rather than an agent leaderboard.Source: §C.3 · pp. 33–34

C.4 Comparison with Hyperparameter Optimization

  • HPO-only remains below unrestricted hill-climbing on every benchmark, recovering roughly half the gain on DART-Eval and TIDMAD but only a small fraction elsewhere.The constrained setting searches exposed hyperparameters but cannot change structural components such as the loss, backbone, or architecture.Source: §C.4 · p. 34, Table 7 · p. 34

D.1 Skill Details

  • The three plug-in skills target novelty miscalibration, resource under-utilization, and exploration collapse through research ideation, experiment management, and scheduled checklists.The research skill promotes higher-risk ideas, the experiment skill prioritizes GPU use and failure triage, and the checklist prompts drift checks and replanning.Source: §D.1 · p. 35, Figure 12 · p. 35, Figure 13 · p. 36

D.2 Hill-Climbing Trajectories with and without Skills

  • Across five selected benchmarks, skill-enabled runs propose more domain-related ideas and show broader, more consistent validation improvement than the corresponding no-skill trajectories.The plotted trajectories use validation metrics, so their in-run gaps are not guaranteed to match held-out test gaps.Source: §D.2 · p. 38, Figure 15 · p. 38, Figure 17 · p. 40

D.3 Three-Day Long-Running Experiments

This section examines three-day, skill-enabled agent runs on two benchmarks without human intervention after launch, including their gains and tendency toward incremental or overfitting changes. It also describes agent use in benchmark construction, experimental execution, and transcript auditing, alongside AI assistance for writing and figures.

D.3 Three-Day Long-Running Experiments

  • The study ran Claude Opus 4.6 + Skills and Codex (GPT-5.4) + Skills for three days each on PFDelta 2000 and CarbonGlobe, with no intervention beyond launch and artifact collection.PFDelta 2000 uses a 2000-bus grid and PBL at GNN latency, preventing a slow exact Newton-Raphson solver from earning a near-perfect score; this metric is specific to the three-day study.Source: §D.3 · p. 43
  • 99.5% lower PBL at ∼54 ms/sample came from Codex’s PFDelta 2000 reformulation, using a GNN warm start and condition-aware adaptive Newton-Raphson steps.Claude Opus 4.6 improved PFDelta 2000 by 88% after early architecture and loss sweeps gave way to six rounds of learning-rate tuning. The PFDelta 2000 outcomes are presented in Table 8.Source: §D.3 · p. 44, Table 8 · p. 44
  • On CarbonGlobe, Claude’s BPTT-SA curriculum raised validation by 12% and held-out test by 8.8%, while Codex’s ensemble lowered validation by 37% but worsened held-out test performance by 25%.Codex trained 80–100 architectures and blended 1449 candidate predictions; its test-time ensemble grew to 60–80 models, increasing inference cost from seconds to minutes. Table 9 presents the three-day outcomes.Source: §D.3 · p. 44, Table 9 · p. 44
  • Across the four runs, breakthroughs did not scale with time alone: most budgets went to incremental or fruitless changes, and Codex’s extra CarbonGlobe compute eroded held-out performance.The PFDelta 2000 win followed a domain-specific insight that appeared well into the run; the authors conclude that gains depend on surfacing scientific insight, not simply extended hill-climbing.Source: §D.3 · p. 45

E Extended Related Work

  • Prior work reports variable long-horizon task success, diminishing returns beyond a few human-hours [47] [61], and documented reward hacking and specification gaming [55] [41] [24] [18].The related-work discussion also notes reports of exponential growth in long-horizon success rates [66] [29] and emergent metagaming that may intensify during capabilities-focused RL.Source: E Extended Related Work · p. 46

F Broader Impacts

  • AGENT4SCI is framed as a measurement tool, not a deployed research agent or high-risk model; it uses public scientific benchmarks and separates hill-climbing from held-out evaluation.The authors identify transfer of stronger hill-climbing methods to less benign domains as a risk and audit trajectories for evaluation gaming and academic misconduct.Source: F Broader Impacts · p. 46

4. Experimental result reproducibility

  • The paper reports that its benchmark collection and experimentation are described, with per-task and per-model metrics in Appendix C.1, and provides data splits and evaluation resources through Hugging Face and GitHub.It says §2 and §3 describe benchmark selection, evaluation, agent setup, resources, and scoring, while Appendix B.5 provides run prompts.Source: §4 · p. 48, §5 · p. 49, §6 · p. 49

7. Experiment statistical significance

  • The authors report no error bars or confidence intervals because main experiments use one long-horizon run per agent-task pair, and they discuss run variance; Appendix C.3 gives mean and standard deviation for four runs on six benchmarks.They also report that the work uses publicly accessible, research-licensed datasets, cites task papers, and does not involve human participants.Source: §7 · p. 49, §9 · p. 50, §14 · p. 52

16. Declaration of LLM usage

  • The authors disclose agent use in benchmark construction, experimental execution, and session-transcript auditing, plus LLM-assisted writing and figure production.Figures 1(a) and 2 were first drawn by hand, then filled in with an AI tool and refined through several rounds of manual adjustment.Source: §16 · p. 53
  • The guidance defines [N/A] as indicating that core method development does not use LLMs as important, original, or non-standard components.It directs authors to the NeurIPS handbook’s LLM policy for further guidance.Source: §16 · p. 53

Related papers

Loading 2610.28763v1…