finding

active

finding:approximately-half-of-the-26-otd-latents-show-near-zero-or-negative-effect-sizes-activating-more-during-on-topic-content

Approximately half of the 26 OTD latents show near-zero or negative effect sizes, activating more during on-topic content

Reveals that contrastive search yields a heterogeneous set, not all functioning as true off-topic detectors

Source paper

extracted_from

Endogenous Resistance to Activation Steering in Language Models

(2026) · Alex McKenzie · Keenan Pepper · Stijn Servaes · Martin Leitgab +5

Neighborhood — ranked by edge-count

Claims (1)

claim

Off-topic detector is a functional label based on selection methodology; these latents may serve broader coherence-monitoring roles beyond detecting off-topic content
supports
Epistemic caution about over-interpreting the OTD label given the heterogeneity of identified latents

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.

OTD latents fire 4.4× higher during off-topic content compared to baseline episodes without self-correctionfinding0.845
Quantitative characterization of OTD activation differential establishing their off-topic monitoring role
OTD latent activation begins declining before verbal self-correction appears in the output in Llama-3.3-70Bfinding0.768
Temporal pattern consistent with internal monitoring process preceding explicit self-correction
Ablating 26 OTD latents reduces multi-attempt rate by 25% (from 7.4% to 5.5%) in Llama-3.3-70Bfinding0.762
Primary causal evidence for dedicated internal consistency-checking circuits
The 26 differentially-activated OTD latents play a causally important role in enabling ESR in Llama-3.3-70Bclaim0.762
Causal interpretation of the ablation experiment results
OTD latent ablation leaves mean first-attempt score unchanged (baseline 26.3, ablation 27.4) in Llama-3.3-70Bfinding0.747
Evidence that OTDs specifically support meta-cognitive monitoring rather than general response generation
Deception feature steering under history, conceptual, and zero-shot controls produces 0% experience reports under both suppression and amplificationfinding0.729
Experiment 2 control analysis confirming gating effect is specific to self-referential processing regime
Backtracking latents remain low during off-topic content and peak shortly after self-correction begins in Llama-3.3-70Bfinding0.727
Complementary temporal activation pattern suggesting distinct roles for OTD and backtracking latent classes
At 100 tokens post-steering, 48 of 171 emotion features remain individually BH-significant despite average effect being near zero.finding0.727
Demonstrates long-tail persistence of causal steering effect in a subset of emotion features