Agent-R1¶
Core mechanism, not full benchmark reproduction.
Paper information¶
| Field | Value |
|---|---|
| Paper | Agent-R1 (arXiv 2511.14460) |
| Affiliations | University of Science and Technology of China |
| First public release | 2025-11-18 (arXiv v1) |
| Authors' code | Author code. |
| Adapter / method | agent-r1 |
| Local code | src/auto_research/agent_research/ |
Original paper¶
Background and contribution¶
Treat each agent/environment interaction as a transition. Pluggable context management, environment interfaces, and optimizers support token- or step-level credit.
flowchart LR
I[Agent state] --> T[Environment transition] --> C[Token or step credit] --> U[Policy update]Key figure from the paper¶
Original Figure 3: Agent-R1. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\[
\tau=(s_t,a_t,o_{t+1})_{t=1}^T,\quad A_t=\delta_t+\gamma\lambda A_{t+1}.
\]
Paper results¶
The technical report supports multi-turn agentic RL workflows. No production A/B.
Local reproduction¶
planbench-mini, 120 episodes, seed 42: joint success 1.0000, average cost 0.9600. See metrics.
auto-research agent-study --method agent-r1 --benchmark planbench-mini --episodes 120 --seed 42
auto-research evolve --model agent --dataset planbench-mini --direction "Combine agent-r1 with installed paper operators" --generations 2 --population 4
Reproduction limits¶
Mechanism/contract diagnostics, not the original model, training scale, environments, or full benchmark.