Steerable but Not Decodable: Function Vectors Operate Beyond the Logit Lens

TL;DR

This paper reveals that function vectors can steer model behaviors without being decodable through traditional logit-based methods, challenging assumptions about linear representations in neural networks.

Contribution

It demonstrates that steerability and decodability are distinct properties, showing FV steering succeeds where decoding fails, with implications for model interpretability and safety.

Findings

FV steering often succeeds without decodable answers at intermediate layers.

Decoding and steering capabilities are largely independent, contradicting previous assumptions.

FV-based interventions can alter model behavior without leaving traceable logit-lens signals.

Abstract

Activation steering presupposes that task-relevant behaviors correspond to linear directions in activation space -- directions that should both steer the model and be readable along the unembedding. Function vectors (FVs), extracted as mean differences across ICL demonstrations, are the canonical test case; the prediction: steering and decoding succeed or fail together. Across 12 tasks, 6 models from 3 families, and 4,032 directed cross-template pairs, we find the opposite. FV steering routinely succeeds where the logit lens cannot decode the correct answer at any intermediate layer, while the converse -- decodable without steerable -- is nearly empty (3 of 72). The gap is not representational dialect. A diagonal tuned lens closes 1 of 14 steerable-not-decodable cases; a 2-layer MLP probe with a Hewitt \& Liang control closes 5 of 10 via nonlinearly encoded structure but leaves 5 invisible to every decoder tested. Even at $> 0.90$ steering accuracy, projecting the FV through the unembedding yields incoherent token distributions: FVs encode computational instructions, not answer directions. A model-family asymmetry sharpens the picture. Mistral FVs rewrite intermediate representations, while Llama and Gemma FVs steer the final output without leaving a logit-lens-visible trace, corroborated by three signals (post-steering deltas, activation-patching recovery, FV norm-transfer correlations). A previously reported negative cosine-transfer correlation dissolves at scale, adding at most $\Delta R^2 = 0.011$ beyond task identity. These results decompose the linear representation hypothesis into linear decodability and linear steerability and show they come apart opposite to intuition, with implications for safety monitoring: vocabulary-projection tools are blind to FV-style interventions on widely deployed model families.