finding
active
finding:steering-qwen-3-8-27b-raises-honeypot-use-from-1-194-to-nearly-a-quarter-contemplating-in-comments-nearly-halfSteering Qwen 3.8 27B raises honeypot use from 1/194 to nearly a quarter, contemplating in comments nearly half
Causal role of probe.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Andrew Lee · Barak Widawsky +14
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Interpretation of logit lens, exemplars, steering.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Smaller models show non-monotonic and diminished ASR with increasing cone dimensionality
- Optimal activation capping layers for Qwen 3 32B are layers 46-53 (out of 64) at 25th percentile capfinding0.738Specific implementation finding for Qwen capping parameters
- Qwen 35B (3B active params, score 4.38) outscores Hermes 405B (405B active params, score 1.75) by 2.5xfinding0.733Parameters don't predict scores; 135x more parameters yields 60% lower score
- Experiment 2 result showing large models can support high-dimensional truth cones
- Quantifies harness activation failure for weak-tier models vs. strong-tier models
- Vulnerability profile for Qwen3.5-27B showing near-zero AS vulnerability
- Qwen3-235B leads as evolver on SWE-bench with 8.2 pp harness-updating gain but ranks last on MCP with 0.6 ppfinding0.721Illustrates benchmark-dependent reshuffling of evolver rankings, no evolver dominates across all substrates
- Optimal steering layers: Llama-3.1-8B layer 12, Gemma-2-9B layer 20, Qwen-2.5-7B layer 14finding0.718Layer sweep result.