Search papers, labs, and topics across Lattice.
This study audits step-level credit assignment methods used to train LLM agents in a controlled environment (ALFWorld) against causal ground truth derived from executed replay. The findings reveal that commonly used credit signals, such as LLM-judge scores and outcome-conditioned logprob ratios, fail to accurately identify causally significant steps, performing no better than chance. The research highlights a critical disconnect between step correctness and step contribution, suggesting that existing evaluation metrics may misrepresent the effectiveness of training signals in LLMs.
Step-level credit assignment methods in LLM training are misleading, with none outperforming random chance in identifying causally significant actions.
Audited against causal ground truth from executed replay in a single-agent tool environment (ALFWorld), none of the step-level credit signals used to train LLM agents -- LLM-judge scores, outcome-conditioned logprob ratios, or the policy's own confidence -- identifies which steps causally matter better than chance. Existing evaluations grade these signals against annotated step *correctness*; we audit them against step *contribution* -- what re-sampling the policy's own alternatives at each decision point and rolling forward actually changes about the outcome -- and the two come apart. The ground truth itself is structured: causal contribution is sparse (30.5% of decision points where ground truth is defined carry measurable effect), and measurability is model-dependent -- the fraction of points with no policy-supported counterfactual differs by a factor of two (13.1% vs. 26.8%) between two similar-scale policies. The failure mode is identifiable: implicit credit echoes the policy's fluency (median rank correlation +0.75, replicating at +0.70 in a second family under a corrected instrument), while conditioning on the outcome adds no causal information (partial correlation -0.004, Qwen). A confidence-only router recovers pivotal steps at chance level, but cuts judge cost by 13.1% per turn (14.0% per trajectory). In a seven-arm pre-registered training experiment, no arm reliably outperforms the untrained policy, and the checkpoints'apparent instrument signature is fully explained by training dose -- sparser credit retains fewer examples, an order-of-magnitude spread in optimizer steps -- not credit content. Comparisons of credit rules must therefore match effective sample size, or they measure dose, not credit.
