Unverified50% confidenceFactExact time
A study from RSS found that on complex reasoning tasks of long-horizon agents, numerical noise accumulates within the verification window, and the draft acceptance results of the thinking process may change the model's output distribution after 30 steps, thereby significantly affecting the reasoning accuracy of long-horizon agents in random directions.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSpline activation functions require table lookup or iterative interpolation at inference time, and their computational cost can be several to dozens of times that of ReLU, requiring careful evaluation in latency-sensitive scenarios such as real-time speech recognition.75% similarUnverifiedA single sample is a very noisy snapshot for models near the passing threshold, and the authors suggest that the next step should be to sample each problem multiple times.74% similarVerifiedInference in diffusion models is an iterative denoising process that typically requires dozens or even hundreds of timesteps and is particularly sensitive to numerical errors.74% similarUnverifiedStep-by-step scoring requires a human or a judge model to read the entire trajectory, which is costly, slow, and noisy.72% similarVerifiedOverly verbose, contradictory, or ambiguous system prompts cause the model's attention to become scattered, resulting in over-thinking or instruction-following failures; streamlining agents.md can improve the determinism of instruction following.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1134610API
curl https://kongchang.com/api/v1/knowledge/claims/1134610MCP
get_claim(id=1134610)