LetTokenBurnMMPostTrainBench: Benchmarking Autonomous Research for Multimodal Post-Training
Evaluates autonomous agents on multimodal post-training across 8 tasks (image, audio, video) measuring model outcomes, iterative improvement, and research integrity.
$ npm i -g lettokenburn-cli$ ltb bench info mmposttrainbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 5 | 44.4 | claude code | Independent | medium | Oct 4, 2026 | arxiv.org |
| GPT-5.6 Sol | 44.4 | codex | Independent | medium | Oct 4, 2026 | arxiv.org |
| Qwen 3.8 Omni Flash | 43.8 | qwen code | Independent | medium | Oct 4, 2026 | arxiv.org |
| GPT-5.6-terra | 43.4 | codex | Independent | medium | Oct 4, 2026 | arxiv.org |
| Gemini 3.8 Flash | 42.9 | gemini cli | Independent | medium | Oct 4, 2026 | arxiv.org |
| Claude Opus 4.8 | 40.6 | claude code | Independent | medium | Oct 4, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
MMPostTrainBench assesses autonomous research agents on eight multimodal tasks spanning image, audio, video, and code repair. Agents operate within fixed budgets to improve a base model using development feedback.
Scoring evaluates target/non-target performance, iterative candidate selection, and research integrity. Results show frequent regressions and selection gaps, with best models trailing top candidates by up to 5.38 percentage points.