Unverified50% confidenceBenchmarkExact time
In the Qwen3-4B comparative experiment, updating the model directly with unfiltered skills without using SGUID filtering decreased performance by 0.3 percentage points on HMMT25.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen using SGUID, Qwen3-4B improved by 0.5 percentage points on HMMT25 after the first round, and by a cumulative 1.1 percentage points after the second round.76% similarUnverifiedThe IQ4XS quantized version of Qwen3 27B shows only approximately 6% performance degradation compared to the original full version.73% similarUnverified3,189 successful rollouts were collected from Qwen3.8-27B for supervised fine-tuning (imitation learning), and performance stagnated at 47.5%, below the results of two rounds of RL training.70% similarUnverifiedExperience internalization fine-tuning improved Qwen3-8B-27B's one-shot score from 37.07 to 48.22, while reducing generated tokens from 6.41 million to 570,000 (an 11-fold reduction).70% similarUnverifiedQwen3.8-Flash-Next surpasses the Qwen3.8-27B dense model in performance on multiple benchmarks while reducing inference and training overhead.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1013308API
curl https://kongchang.com/api/v1/knowledge/claims/1013308MCP
get_claim(id=1013308)