Unverified50% confidenceBenchmarkExact time
On five public math benchmarks, compared with the same training steps and data volume, CoRD improves average pass@1 by 0.9 to 2.0 percentage points and pass@8 by 1.2 to 3.4 percentage points over online distillation.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/6/2026
First Seen
Valid until: 2/4/2027
Sources
Related Entities
Related Claims
UnverifiedAfter RL post-training, the small model's accuracy improves from 4% to 45% (pass@1).73% similarUnverifiedIncreasing the number of online samples per prompt from 1 to 7 changes the LC win rate by at most 2.1 points; sweeping the margin from 1 to 5 changes the LC win rate by at most 2.4 points in the independent pipeline.69% similarUnverifiedIn a benchmark on 50 synthetic chore scenarios, the fine-tuned SlothDistill, compared with the base Llama-3-8B, reduced p95 decision latency from 1,420 ms to 185 ms (a 7.6x improvement), cut average output tokens from 412 to 48 (an 88.3% saving), raised direct-action accuracy from 62.4% to 97.8%, and lowered API cost per thousand tasks from $1.85 to $0.22 (88.1% cheaper).66% similarUnverifiedOn the hard tier, recall@5 improved from 0.000 to 0.955 (under default blending weights), and reached 1.000 when using the lower-bound variant.65% similarUnverifiedThe transfer experiments in the study show that fine-tuning on the same dataset yields performance gains ranging from +6.8 to +23.1 percentage points.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1141700API
curl https://kongchang.com/api/v1/knowledge/claims/1141700MCP
get_claim(id=1141700)