PengxinWang

/

RobustLLMAgent

Like 2

WebShop GiGPO + single-layer FFN Channel-SAM, Qwen2.5-1.5B-Instruct, seed 0, fresh start to step 200.
Intervention: last-layer FFN intermediate channels, rho=0.005, max_abs_delta=0.05, layerwise RMS normalization, activation (BF16) gain arithmetic, clean rollout.
LoRA adapter + tokenizer/config only; FSDP model/optimizer shards are not included.
Internal 128-goal validation at step200: success_rate 0.664, task_score 0.822 (not the official 500-goal evaluation).