finding
active
finding:benchmark-identification-rate-rises-from-0-08-sft-to-0-25-dpo-interventions-barely-reduce-itBenchmark identification rate rises from 0.08 (SFT) to 0.25 (DPO); interventions barely reduce it
Table 11.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Sidharth Baskaran · Max Loeffler +13
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Prompt-conditioned behaviors are entangled with broader response concepts and not cleanly separablesupportsExplains partial recovery.
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.
- Physics sycophancy user-praise rises from 1.32 (SFT) to 2.18 (DPO); reward shaping α=4 reaches 1.80finding0.790Table 11.
- Main evaluation result showing best variant outperforms many proprietary and open-source baselines of comparable or larger sizes.
- SFT on correction+retraction data gives 39.4% inline overall reduction and 9.6% not-inline, well below RLFRfinding0.773SFT baseline comparison.
- SFT baseline failure.
- DPO on Dolci yields lower aggregate safety score than SFT checkpoint (e.g. Llama 3.1 8B 0.849 to 0.758)finding0.768Safeguard degradation.
- Table 11.
- Fig. 12.
- Test-time probe metrics.