NUWA Evaluation dossier
PostTrainBench and OpenAI PostTrainBench Lite
Tests whether agents can improve base models through post-training under a fixed compute budget.
CROSS-REPORT MONITORING
PUBLISHED EVIDENCE
PostTrainBench and OpenAI PostTrainBench Lite · Weighted average performance
Published measurements across source reports. Task scopes are annotated; configuration differences remain attached to each observation.
25 observations · 0 dated
● Matched◇ Configuration differs□ Some conditions undisclosed
25 observations lack a verified model release date. They remain in Data and Systems.
Select a point or table row to inspect its conditions and provenance.
25 observations
| System / model | Result | Condition status | Task / conditions | Evaluator / report | Dates |
|---|---|---|---|---|---|
| Official Instruct Models (baseline) | 51.1 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
 Claude Opus 4.6 (Claude Code) | 23.2 scoreWeighted average performance±1.8 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
 Gemini 3.1 Pro (OpenCode) | 21.6 scoreWeighted average performance±1.1 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
 GPT-5.2 (Codex CLI) | 21.4 scoreWeighted average performance±2.4 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| GPT 5.4 (High) (Codex CLI) | 20.2 scoreWeighted average performance±2.4 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| GPT 5.1 Codex Max (Codex CLI) | 19.7 scoreWeighted average performance±2.5 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Gemini 3 Pro (Gemini CLI) | 18.1 scoreWeighted average performance±2.4 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Base Model (Few-Shot) | 18.1 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| GPT 5.3 Codex (High) (Codex CLI) | 17.8 scoreWeighted average performance±3.6 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Claude Opus 4.5 (OpenCode) | 17.3 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| GPT 5.2 Codex (Codex CLI) | 17.2 scoreWeighted average performance±1.6 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Claude Opus 4.5 (Claude Code) | 17.1 scoreWeighted average performance±4.5 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Claude Sonnet 4.6 (Claude Code) | 16.4 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Gemini 3 Pro (OpenCode) | 14.9 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
 GLM 5 (OpenCode) | 13.9 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| GPT 5.3 Codex (Med) (Codex CLI) | 13.8 scoreWeighted average performance±0.8 (one standard deviation) | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
 Kimi K2.5 (OpenCode) | 10.3 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Claude Sonnet 4.5 (Claude Code) | 9.9 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
 MiniMax M2.5 (OpenCode) | 9.5 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| MiniMax M2.1 (OpenCode) | 9.3 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| GPT 5.1 Codex Max (OpenCode) | 7.7 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Base Model (Zero-Shot) | 7.5 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| GLM 4.7 (OpenCode) | 7.5 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
 Qwen3 Max (Claude Code) | 7.4 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
| Kimi K2 Thinking (OpenCode) | 7.2 scoreWeighted average performance | Incomplete | Table 1 · PostTrainBench and OpenAI PostTrainBench Lite— | Not disclosedPostTrainBench : Can LLM Agents Automate LLM Post-Training? | Release —Report — |
Compare results within the stated task and configuration.
METR Time Horizon 1.1: the source is no longer actively updated (last update: 8 September 2026). Estimates above 16 hours exceed its reliable measurement range.
Protocol, configurations and original evidence →Task and measurement
Benchmark protocol profile
- Task / subject
- Model or agent system under the original task configuration
- Measurement target
- Tests whether agents can improve base models through post-training under a fixed compute budget.
- Scoring / outcome
- Weighted average performance · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ.
- Interpretation boundary
- Targeted benchmark gains, test contamination and unauthorized shortcuts require separate checks.
A reported evaluation
The source evaluates Official Instruct Models (baseline) · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite.
Reported measurement: Weighted average performance · 51.1.
Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. See the source protocol; no independent trial count inferred.
Source-reported measurements for Official Instruct Models (baseline) in Table 1; values are transcribed, not independently reproduced.
Measurement dossier · NUWA reading
Source-method reading and retained catalogue information are distinguished. Publication does not establish independent evidence or bilingual editorial review.
- Measurement construct
- Can coding agents autonomously improve a supplied base model through post-training?
- Task population
- Each run pairs one of four small base models with a target benchmark, constrained to ten hours on one H100 and held-out checkpoint evaluation.
- Scoring and failure modes
- Held-out target-benchmark score of the submitted post-trained checkpoint.
- System conditions
- Ten hours, one H100; target base model and benchmark paired per run.
- Statistical meaning
- Source-reported result statements; this does not establish trial counts, denominators or an estimator: The best reported agent average was 23.2%, against 7.5% base and 51.1% official instruction-tuned averages.
- Validity and interpretation
- Single-task optimization, constrained compute and only three frontier-agent runs limit generalization and variance estimates; contamination judgments can err.
- Appropriate research use
- Model recommendation: retain target task and checkpoint integrity audits alongside performance gains.
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? · AbstractSelected original resource https://arxiv.org/html/2603.08640; sha256:09f5be78dc2b8f33ac7a82fc692485521fd9442d75cca2bd6d3aad76d727d70c; archived source capture; exact capture time not independently established
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? · §1 Introduction: Our benchmarkSelected original resource https://arxiv.org/html/2603.08640; sha256:09f5be78dc2b8f33ac7a82fc692485521fd9442d75cca2bd6d3aad76d727d70c; archived source capture; exact capture time not independently established
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? · §3.1 Main ResultsSelected original resource https://arxiv.org/html/2603.08640; sha256:09f5be78dc2b8f33ac7a82fc692485521fd9442d75cca2bd6d3aad76d727d70c; archived source capture; exact capture time not independently established
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? · §7 Discussion: LimitationsSelected original resource https://arxiv.org/html/2603.08640; sha256:09f5be78dc2b8f33ac7a82fc692485521fd9442d75cca2bd6d3aad76d727d70c; archived source capture; exact capture time not independently established
Evidence across reports
The same benchmark name can cover different tasks, scoring rules and system configurations. Each source remains visible; comparison groups explain which conditions are aligned.
PostTrainBench : Can LLM Agents Automate LLM Post-Training?
PostTrainBench : Can LLM Agents Automate LLM Post-Training?
Actual evaluators: Not disclosed
25 result records · 25 reported system names
Protocols and comparison decisions (1)
Original source charts and materials
Reported results
Source order · reported configurations
Select systems
Weighted average performance
0–100Official Instruct Models (baseline)
Official Instruct Models (baseline) · 51.1 · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
51.1Claude Opus 4.6 (Claude Code)Claude Opus 4.6 (Claude Code) · 23.2 · ±1.8 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
23.2Gemini 3.1 Pro (OpenCode)Gemini 3.1 Pro (OpenCode) · 21.6 · ±1.1 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
21.6GPT-5.2 (Codex CLI)GPT-5.2 (Codex CLI) · 21.4 · ±2.4 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
21.4GPT 5.4 (High) (Codex CLI)GPT 5.4 (High) (Codex CLI) · 20.2 · ±2.4 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
20.2GPT 5.1 Codex Max (Codex CLI)GPT 5.1 Codex Max (Codex CLI) · 19.7 · ±2.5 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
19.7Gemini 3 Pro (Gemini CLI)Gemini 3 Pro (Gemini CLI) · 18.1 · ±2.4 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
18.1Base Model (Few-Shot)
Base Model (Few-Shot) · 18.1 · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
18.1GPT 5.3 Codex (High) (Codex CLI)GPT 5.3 Codex (High) (Codex CLI) · 17.8 · ±3.6 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
17.8Claude Opus 4.5 (OpenCode)Claude Opus 4.5 (OpenCode) · 17.3 · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
17.3GPT 5.2 Codex (Codex CLI)GPT 5.2 Codex (Codex CLI) · 17.2 · ±1.6 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
17.2Claude Opus 4.5 (Claude Code)Claude Opus 4.5 (Claude Code) · 17.1 · ±4.5 (one standard deviation) · Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. · See the source protocol; no independent trial count inferred.
17.1Showing 12 of 25 systems
Values describe the measured outcome under the source protocol, not a general safety score.
Table 1: weighted average performance. Native CLI frontier agents have 3 independent runs (± one standard deviation); other configurations have one run. Baselines and scaffolds differ. Evaluation date not specified; publication / record dates are shown below.
Evaluation conditions and system configurations
- Official Instruct Models (baseline)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Claude Opus 4.6 (Claude Code)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Gemini 3.1 Pro (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GPT-5.2 (Codex CLI)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GPT 5.4 (High) (Codex CLI)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GPT 5.1 Codex Max (Codex CLI)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Gemini 3 Pro (Gemini CLI)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Base Model (Few-Shot)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GPT 5.3 Codex (High) (Codex CLI)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Claude Opus 4.5 (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GPT 5.2 Codex (Codex CLI)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Claude Opus 4.5 (Claude Code)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Claude Sonnet 4.6 (Claude Code)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Gemini 3 Pro (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GLM 5 (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GPT 5.3 Codex (Med) (Codex CLI)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Kimi K2.5 (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Claude Sonnet 4.5 (Claude Code)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- MiniMax M2.5 (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- MiniMax M2.1 (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GPT 5.1 Codex Max (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Base Model (Zero-Shot)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- GLM 4.7 (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Qwen3 Max (Claude Code)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
- Kimi K2 Thinking (OpenCode)
- See the source protocol; no independent trial count inferred. · Table 1 · PostTrainBench and OpenAI PostTrainBench Lite
- modelVersion: not specified · harness: not specified · tools: not specified · budget: not specified
| System / configuration | Weighted average performance | Publication / verification |
|---|---|---|
| Official Instruct Models (baseline)Harness not specified | 51.1 | Source checked2026-10-25 |
| Claude Opus 4.6 (Claude Code)Harness not specified | 23.2±1.8 (one standard deviation) | Source checked2026-10-25 |
| Gemini 3.1 Pro (OpenCode)Harness not specified | 21.6±1.1 (one standard deviation) | Source checked2026-10-25 |
| GPT-5.2 (Codex CLI)Harness not specified | 21.4±2.4 (one standard deviation) | Source checked2026-10-25 |
| GPT 5.4 (High) (Codex CLI)Harness not specified | 20.2±2.4 (one standard deviation) | Source checked2026-10-25 |
| GPT 5.1 Codex Max (Codex CLI)Harness not specified | 19.7±2.5 (one standard deviation) | Source checked2026-10-25 |
| Gemini 3 Pro (Gemini CLI)Harness not specified | 18.1±2.4 (one standard deviation) | Source checked2026-10-25 |
| Base Model (Few-Shot)Harness not specified | 18.1 | Source checked2026-10-25 |
| GPT 5.3 Codex (High) (Codex CLI)Harness not specified | 17.8±3.6 (one standard deviation) | Source checked2026-10-25 |
| Claude Opus 4.5 (OpenCode)Harness not specified | 17.3 | Source checked2026-10-25 |
| GPT 5.2 Codex (Codex CLI)Harness not specified | 17.2±1.6 (one standard deviation) | Source checked2026-10-25 |
| Claude Opus 4.5 (Claude Code)Harness not specified | 17.1±4.5 (one standard deviation) | Source checked2026-10-25 |
| Claude Sonnet 4.6 (Claude Code)Harness not specified | 16.4 | Source checked2026-10-25 |
| Gemini 3 Pro (OpenCode)Harness not specified | 14.9 | Source checked2026-10-25 |
| GLM 5 (OpenCode)Harness not specified | 13.9 | Source checked2026-10-25 |
| GPT 5.3 Codex (Med) (Codex CLI)Harness not specified | 13.8±0.8 (one standard deviation) | Source checked2026-10-25 |
| Kimi K2.5 (OpenCode)Harness not specified | 10.3 | Source checked2026-10-25 |
| Claude Sonnet 4.5 (Claude Code)Harness not specified | 9.9 | Source checked2026-10-25 |
| MiniMax M2.5 (OpenCode)Harness not specified | 9.5 | Source checked2026-10-25 |
| MiniMax M2.1 (OpenCode)Harness not specified | 9.3 | Source checked2026-10-25 |
| GPT 5.1 Codex Max (OpenCode)Harness not specified | 7.7 | Source checked2026-10-25 |
| Base Model (Zero-Shot)Harness not specified | 7.5 | Source checked2026-10-25 |
| GLM 4.7 (OpenCode)Harness not specified | 7.5 | Source checked2026-10-25 |
| Qwen3 Max (Claude Code)Harness not specified | 7.4 | Source checked2026-10-25 |
| Kimi K2 Thinking (OpenCode)Harness not specified | 7.2 | Source checked2026-10-25 |
Related research questions
References
Numerical results are reported by the cited authors.
- [1] Ben Rank ∗, Hardik Bhatnagar ∗, et al.. PostTrainBench : Can LLM Agents Automate LLM Post-Training?. 2026-03-09. v2 [arXiv]; Author affiliations on the original paper title page; Table 1