IC-168Control vectors derived from BABI improve GSM8K accuracy and vice versa on Mistral-7B-Instruct, indicating a task-general reasoning direction in the residual stream

Bertram Højer, Oliver Simon Jarvis, Stefan Heinrich

SourceImproving Reasoning Performance in Large Language Models via Representation Engineering

The authors derive a PCA control vector from BABI contrastive pairs and apply it to the GSM8K evaluation, and vice versa. In both directions, Mistral-7B-Instruct shows an increase in logit-based accuracy comparable to the same-task control vector, with similar trends in KL divergence and entropy. This cross-task transfer suggests that the residual stream encodes a direction associated with reasoning more broadly rather than a task-specific feature.

Evidence
interventional
Caveat
The authors note that the model representations are not very robust to the intervention (jagged trend line) and that the cross-task result is observed on a single model (Mistral-7B-Instruct) with a single pair of tasks.
Model
Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral-7B-Instruct
Concepts
Linear representation
Datasets
bAbI [eval], GSM8K [eval]
Methods
Representation Engineering / Representation engineering (control vectors) / Zou et al. 2023 (Representation Engineering) / Zou et al. (representation engineering) [primary], Principal component analysis [primary]
Related work
Zou et al. 2023 (Representation Engineering) [builds-on]
Related findings
IC-167
Extraction
automatic-extraction