CAMEL¶
Core mechanism, not full benchmark reproduction.
Paper information¶
| Field | Value |
|---|---|
| Paper | CAMEL (arXiv 2303.17760) |
| Affiliations | King Abdullah University of Science and Technology |
| First public release | 2023-03-31 (arXiv v1) |
| Authors' code | Author code. |
| Adapter / method | camel |
| Local code | src/auto_research/agent_research/ |
Original paper¶
Background and contribution¶
Inception prompts define user and assistant roles, goals, and constraints. Alternating messages solve tasks and produce multi-agent trajectories for study.
flowchart LR
R[Role and goal setup] --> A[Assistant message] --> U[User response] --> AKey figure from the paper¶
Original Figure 1: CAMEL. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
\[
m_t^A\sim\pi_A(\cdot|r_A,g,h_t),\quad m_t^B\sim\pi_B(\cdot|r_B,g,h_t,m_t^A).
\]
Paper results¶
NeurIPS 2023 studies instruction-following cooperation. No production A/B.
Local reproduction¶
planbench-mini, 120 episodes, seed 42: joint success 1.0000, average cost 1.3500. See metrics.
auto-research agent-study --method camel --benchmark planbench-mini --episodes 120 --seed 42
auto-research evolve --model agent --dataset planbench-mini --direction "Combine camel with installed paper operators" --generations 2 --population 4
Reproduction limits¶
Deterministic mini-suite, not a large-model, multi-GPU, private-environment, or full benchmark reproduction.