Dream-1 world action model scores 159% higher after fine-tuning, beating frontier LLMs on real robots
SE3 Labs released an evaluation series, "A Wider Dream," systematically comparing frontier LLM zero-shot capabilities against the fine-tuned world action model Dream-1 across 10 real-robot tasks. The conclusions cut both ways: LLMs still lead overall in zero-shot settings, but after targeted fine-tuning, Dream-1 overtakes them on physical routine tasks at lower cost and in less time. The report and dataset are open-sourced—worth a look for anyone tracking robot foundation-model approaches.
Confirmed
- In the zero-shot stage, Fable (5.2) led pre-fine-tuning Dream-1 on about 7 of 10 tasks; Dream-1 won only on 3 physical routines, such as packing (54.5 vs. 39.0) and toy stacking (35.0 vs. 31.0).
- Memory and reasoning tasks are LLM strongholds: on hidden-cup find-the-block (flipping cups in order), Fable scored 64 versus Dream-1's 9; on sorting objects into three bins by category, Fable 62.5 versus Dream-1 10.
- SE3 Labs fine-tuned Dream-1 with 500 real robot demonstrations, lifting its RoboDojo-Real benchmark score from 8.5 to 22, a 159% gain that surpasses fine-tuned π0.5 (20.6); on repetitive physical routines it beat frontier LLMs in about one-fifth the time.
- The key to the gain was not a bigger model but randomized demonstration data: Dream's zero-shot skills had clustered in the mid-table area where teleoperation data was dense; after switching to demonstrations of 10 tasks with randomized full-table layouts, skill coverage spread across the whole table.
- Fine-tuned Dream-1 still lags on some tasks; cost is reportedly 1/100 that of competitors per the authors; the 500 randomized demonstrations dataset and evaluation report are open-sourced.
Why It Matters
- The evaluation points to a pragmatic path: rather than chasing larger models, fine-tune a world action model with a small number of randomized real-robot demonstrations to approach or even surpass frontier LLMs on physical tasks, with markedly lower data and cost barriers.
2026-11-19 ~ 2026-11-19 · 6 related posts
Primary sources
- Fine-tuned Dream-1 world action model jumps 159% on real-robot benchmark at 1/100th the cost — paigeinsf · 2026-11-19
- [source] Fine-tuned World Action Model Dream-1 Jumps 159%, Beats Frontier LLMs on Robot Tasks — paigeinsf · 2026-11-19
- The Secret Wasn't a Bigger Model: 500 Randomized Demos Lift Dream-1 Score 159% — paigeinsf · 2026-11-19
- Memory and Reasoning Remain LLM Territory: Fable Beats Dream-1 64 to 9 on Cup Task — paigeinsf · 2026-11-19
- [source] LLMs zero-shot beat fine-tuned world action models on 7 of 10 robot tasks — paigeinsf · 2026-11-19
- [source] SE3 Labs Releases A Wider Dream Eval, Opens 500-Episode Robot Demo Dataset on Hugging Face — paigeinsf · 2026-11-19