Unverified50% confidenceBenchmarkExact time
StopWatch achieves state-of-the-art results on six benchmarks across three scene sets through explicit object grounding (locating the target, judging visibility, and estimating distance).
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/19/2026
First Seen
Valid until: 1/17/2027
Sources
Related Entities
Related Claims
UnverifiedOn the clean benchmark, when ground-truth object grounding information is provided, StopWatch's policy performance can nearly match the expert policy that generated its demonstration data.76% similarUnverifiedStopWatch corrects annotation flaws in existing benchmarks, making evaluation more trustworthy.71% similarUnverifiedGoal模式在每一次工具调用之前都会做校验，通过工具调用拦截层将调用意图与预设约束比对，从而杜绝越界操作65% similarUnverifiedThree checks carry most of the evaluation work: comparing actual step counts against expected ranges, where runs that hit the step limit are cleanly separated from runs that finish in four steps; checking tool calls, where tools that were called but returned no usable result, and tools that should have been triggered but never were, have nowhere to hide; and checking the trajectory, i.e. the action sequence, which catches reruns, skipped steps and loops.65% similarUnverified跑分排名不等于所有场景的实际表现，具体任务仍需具体测试62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1065828API
curl https://kongchang.com/api/v1/knowledge/claims/1065828MCP
get_claim(id=1065828)