finding

active

finding:steering-vector-based-instruction-discovery-outperforms-input-embedding-similarity-baseline-for-reflection-inducing-instruction-selection

Steering vector-based instruction discovery outperforms input embedding similarity baseline for reflection-inducing instruction selection

Demonstrates that surface-level embedding similarity fails to capture reflective semantics.

Source paper

extracted_from

Unveiling the Latent Directions of Reflection in Large Language Models

(2025) · Chang, Fu-Chieh · Lee, Yu-Ting · Wu, Pei-Yuan

Neighborhood — ranked by edge-count

Claims (2)

claim

Steering vectors enable systematic discovery of reflection-inducing instructions beyond trial-and-error prompt design.
supports
Core applied contribution claim, supported by top-k accuracy comparisons.
Steering vectors capture latent dimensions of reflective behavior more faithfully than surface-level embedding similarity.
supports
Supported by the instruction discovery experiments comparing steering vs. embedding baselines.

Questions (1)

question

How can we systematically identify effective reflection trigger instructions, rather than relying on trial-and-error?
answered_by
First key research question motivating the methodology.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Steering vectors from µ(0→2) slightly outperform µ(1→2) for instruction discovery across datasets and modelsfinding0.830
Shows that contrasting No Reflection with Triggered Reflection provides a stronger signal than Intrinsic vs Triggered.
Steering vectors discover effective triggers such as 'However' and 'Otherwise', consistent with prior reported reflection datasetsfinding0.804
Validates that steering vectors capture reflection semantics by finding tokens reported in related work.
Steering vector extracted from final post-expert-iteration model also successfully elicits deployment behaviorfinding0.803
Replicates main result using in-distribution steering vector; addresses concern about pre-trained vector validity.
Input embedding similarity baseline selects semantically related but non-reflective tokens (e.g., Await, ConfigureAwait, Unchecked) that fail to improve accuracyfinding0.801
Demonstrates the failure mode of surface-level similarity for instruction discovery.
Steering vector constructed from all 16 contrastive pairs outperforms most single-pair vectors; best 4-pair vector outperforms full 16-pair vectorfinding0.783
Demonstrates averaging multiple prompt pairs reduces noise; optimal subset selection further improves performance.
Input Embedding Similarity Baselinemethod0.779
Baseline method for instruction discovery using surface-level input embedding similarity instead of steering vectors.
Under steering vector interventions, the model relaxes its ethical standards and interprets neutral prompts as implicit suggestions to deceive, creating ethical dilemmas triggering repetitive reasoning cyclesclaim0.775
Mechanistic interpretation of how activation steering induces deception through the model's reasoning process
Cosine Similarity Ranking for Instruction Discoverymethod0.773
Method to discover new reflection-inducing instructions by ranking candidate tokens by cosine similarity to steering vectors.