Unverified50% confidenceBenchmarkExact time
Using only synthetic weak instructions fed directly into the navigator, the success rate was 15.27%, falling between the exhaustive command rate of 31.03% and the real user instruction rate of 11.33%.
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAfter TGIT training, the success rate for weak instruction input increased from 15.27% to 37.93%, surpassing the original exhaustive command rate of 31.03%.75% similarUnverifiedIn the configuration using ten policy snapshots, using only 8 prompts achieved an average accuracy of 24.09%, approaching the 24.51% achieved with 14,080 prompts.69% similarUnverifiedFor comparison, the Program-of-Thought baseline achieves 72.0% overall accuracy and the tool-calling agent achieves 58.7%.69% similarUnverifiedOn a test subset of 152 completely untouched questions, anchored hybrid retrieval achieved 42.11%, while the planner-guided method reached only 36.84%.66% similarUnverifiedWith the prompt unchanged, simply adjusting the thin direction encoding "who is endorsing" can shift compliance by 11 to 32 percentage points and close 55% to 61% of the gap between source and user.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990449API
curl https://kongchang.com/api/v1/knowledge/claims/990449MCP
get_claim(id=990449)