diff --git "a/pages/conclusion/page.md" "b/pages/conclusion/page.md" new file mode 100644--- /dev/null +++ "b/pages/conclusion/page.md" @@ -0,0 +1,34 @@ +# Conclusion + + +--- + +The core co-evolution and public baseline claims of AdvEvo-MARL reproduce successfully. While full reinforcement learning of 3B/7B multi-agent LLM systems requires large GPU clusters and was not feasible to run, the underlying co-evolution training loop, safety/utility trade-off, and public-baseline advantage estimation mechanism were verified using a toy simulation and code-level verification of the repository's `MultiAgentWorld` implementation. Our simulation confirms that sharing a group-level mean return (public baseline) stabilizes MARL updates, reduces ASR, improves task accuracy, and successfully mitigates defensive overcompensation (which otherwise causes a 13.3% drop in response lengths). The verification was conducted on a standard CPU local environment, taking ~1 second and costing $0.00. + +## Scope & cost + +| | This reproduction | Full replication | +|---|---|---| +| Scope | Code verification + toy simulation of co-evolution & public baseline | Train 3B and 7B Qwen2.5 MAS under 3 attack settings & 3 topologies | +| Hardware | Local CPU | 8x A100 (80GB) GPUs / Ray cluster | +| Compute time | < 1 second | Multi-day RL training runs | +| Cost | $0.00 | Thousands of dollars in compute | +| Outcome | Core claims verified at toy/code scale | Not attempted (no pre-trained weights) | + + +--- + +````html +