The authors train a 9B-parameter multimodal agent using online reinforcement learning (RL) with a 128k-token context window and up to 75 tool-interaction turns per episode.
Training long-horizon multimodal research agents with reinforcement learning at scale
Long-MDR introduces a three-component recipe enabling stable online RL training with 128k context and 75+ tool-interaction turns.
Independent
On Tai Tang
Research Digest··3 min read
On Tai Tang studies online reinforcement learning for multimodal deep-research agents at unprecedented scale: 128k context and 75+ tool-interaction turns.
Why this paper
Independent
In one line
Long-MDR makes 128k-context, 75-turn multimodal reinforcement learning more efficient and stable through distillation warmup, gradual horizon expansion, and entropy-based recovery.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§