finding
active
finding:mildly-poisoning-5-of-sft-and-dpo-data-induces-each-target-trait-interventions-substantially-reduce-trait-expressionMildly poisoning 5% of SFT and DPO data induces each target trait; interventions substantially reduce trait expression
Fig. 3 on Llama-3.1-8B.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Unifying Bayesian lens.
- Poisoning results.
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.
- Fig. 12.
- Benchmark identification rate rises from 0.08 (SFT) to 0.25 (DPO); interventions barely reduce itfinding0.757Table 11.
- Confirms attention rather than MLP as the locus of persona generation
- Shows SFT effect is trait-specific and reflects register of demonstrations
- Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance
- SAE feature monitoring detects misalignment risk before behavioral evaluation can
- Fig. 10, Table 10.
- Key claim about the value of the introspection stage, supported by both prefill attack and adversarial prompting experiments