IC-930StackLLama, when used as a reward model, achieves near-random consistency on contrast instructions for the Stack Exchange task

Lingfeng Shen, Sihao Chen, Linfeng Song, Lifeng Jin, Baolin Peng, Haitao Mi, Daniel Khashabi, Dong Yu

SourceThe Trickle-down Impact of Reward Inconsistency on RLHF

The paper evaluates the released StackLLama checkpoint on contrast instructions, a benchmark of lexically similar instruction pairs with different ground-truth responses. A consistent reward model should assign higher reward to the correct instruction-response pairing than to the swapped distractor. StackLLama scores 50.1 on response consistency (cres) and 48.2 on instruction consistency (cins) for the Stack dataset, essentially at the random-guessing baseline of 50.0. This is in stark contrast to human annotators who achieve 81.0 (cres) and 82.1 (cins) on the same dataset, indicating a large gap between the model's reward assignments and human preference judgments.

Evidence
correlational
Key metric
cres 50.1, cins 48.2 (StackLLama, Stack contrast instructions) vs. random 50.0 and human 81.0 / 82.1
Caveat
The evaluation uses the StackLLama policy checkpoint as a reward model, which is not its intended use; the near-random result may reflect the mismatch between a policy model and the reward-model task rather than a fundamental limitation of the model.
Model
StackLLaMA
Concepts
Failure mode
Datasets
StackExchange [source]
Related work
StackLLaMA [context]
Related findings
IC-931
Extraction
automatic-extraction