Unverified50% confidenceFactExact time
Un-tuned PPI with fixed λ=1 inflates human-label consumption by a factor of 2.27, which amounts to injecting noise in medium-quality settings where judge agreement is only 45%–77%.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAblation analysis shows that fine-tuning on the segment paradigm yields a +5.29-point improvement on the test set, while replacing the first 4,500 transcript words with 2,000 retrieved words yields only a +1.55-point improvement.68% similarUnverifiedMeasured on 45 real candidates, the similarity between the HRR query vector and the stored fact vectors fluctuated between -0.023 and +0.015, with a span of 0.038 constituting noise; after rescaling, the term's contribution to each score was 0.489–0.508, i.e., a near-constant 0.147–0.152 in the final result.67% similarUnverifiedUnder the most permissive setting of each relaxed verification rule, relaxed verification accepted an average of 1.34 times as many tokens per round as lossless verification, but the generated completion length inflated to 1.54 times.66% similarUnverifiedCompared with the improvement under the single-appearance condition, the paired improvement of Concat IDM under the multi-appearance condition was 52.9 percentage points higher.64% similarUnverifiedThe GPQA Diamond subset has 198 questions in total, and at a 95% confidence interval the standard error of a single sample is about 2 percentage points, so a difference of 0.7 points is within the noise range.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1069906API
curl https://kongchang.com/api/v1/knowledge/claims/1069906MCP
get_claim(id=1069906)