Unverified50% confidenceBenchmarkExact time
In language modeling experiments, using the generation perplexity of Qwen 2.5 1.5B as the oracle reward and BERT, distilled BERT, token entropy, etc. as surrogates, with a feedback set of only 256 values or rankings, the KL between the q_G learned by Martini and q* is significantly lower than that of uniform weighting or Pareto-type methods.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOn the GSM8K mathematical reasoning task, Qwen model adaptation can maintain answer format compliance, but the exact-match retention effect is confounded by random seeds at the 0.6B scale and even deteriorates at the 1.7B scale.69% similarUnverifiedChoosing a model at the 4B parameter level (such as Qwen3.5-4B) is a pragmatic engineering choice: neuron-by-neuron visualization and editing on larger models is unrealistic in compute and memory, while smaller models have semantic behavior that is not rich enough.67% similarUnverifiedThe author recommends Q4_K_M or Q5_K_M quantization for code assistance, and Q5_K_M or Q6_K for math and logic tasks to ensure reasoning quality.66% similarUnverifiedPure reasoning models tend to produce unstable formatting when generating structured tool-calling instructions, and Qwen3's hybrid training has resolved this issue.66% similarUnverifiedTesting on Qwen2.5-7B, Llama-3.1-8B, and two DeepSeek models found that selecting memory by relevance reintroduces decoy facts.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1127430API
curl https://kongchang.com/api/v1/knowledge/claims/1127430MCP
get_claim(id=1127430)