IC-219The bag-of-heuristics mechanism in LLaMA3-8B fails on certain arithmetic prompts due to insufficient total logit contribution from heuristic neurons, not due to a lack of associated heuristics

Yaniv Nikankin, Anja Reusch, Aaron Mueller, Yonatan Belinkov

SourceArithmetic Without Algorithms: Language Models Solve Math with a Bag of Heuristics

The authors test two hypotheses for why the bag of heuristics produces incorrect answers: (i) the prompt lacks sufficient associated heuristic neurons, or (ii) the heuristics have imperfect recall or low logits for the correct answer token. Sampling 50 correct and 50 incorrect prompts, they find that incorrect prompts actually have MORE associated heuristic neurons on average, ruling out hypothesis (i). However, the total logit contribution of all heuristic neurons to the correct answer token is slightly higher for correct prompts than for incorrect ones, supporting hypothesis (ii): the primary failure mode is poor promotion of the correct answer logit, not an absence of heuristics.

Evidence
correlational
Key metric
incorrect prompts have more associated heuristic neurons than correct prompts on average; slight advantage in total logit contribution for correct prompts over incorrect prompts
Caveat
Analysis is limited to 50 correct and 50 incorrect prompts sampled from LLaMA3-8B; the logit contribution difference is described as 'slight' and the exact magnitude is shown only in a figure.
Model
Llama 3 8B
Concepts
Failure mode
Methods
Logit lens [primary]
Related findings
IC-218, IC-220
Extraction
automatic-extraction