PengxinWang

/

RobustLLMAgent

Like 2

ALFWorld GiGPO + single-layer FFN Channel-SAM, Qwen2.5-1.5B-Instruct, seed 0, fresh start to step 200.
Intervention: last-layer FFN intermediate channels (8960 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.
Training-time validation on the complete fixed splits (valid_seen 140 / valid_unseen 134), not the official merged-vLLM evaluation protocol:
step100: ID 0.807, OOD 0.828
step200: ID 0.907, OOD 0.866
Run settings are the launcher defaults for this benchmark (tensor_model_parallel_size=1, load_format=dummy_dtensor, actor micro-batch 8, no parameter offload); see run_config.json.