AI Paper Atlas

Research map / Large language models / Agents, code and math

PILOT-Bench: Probabilistic Interaction for LLM Operations in Tool-driven Scenarios

Ruicheng Ao, Zeping Min, Tingyu Zhu, Wotao Yin, Xinshang Wang

ICLR 2026 · Large language models › Agents, code and math · approach: Software engineering agents

This paper tried a common change: Chain-of-thought prompting

add Chain-of-Thought instructions to baseline prompt · ResiliBench simulated · full success rate (%): 51.4 -> 50.8 No clear effect

What 1,107 accepted papers report for this change

See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace baseline prompt with MDP-optimal workflow prompt7 LLMs average · ResiliBench simulatedfull success rate (%): 51.4 -> 62.1Improved
add Chain-of-Thought instructions to baseline prompt7 LLMs average · ResiliBench simulatedfull success rate (%): 51.4 -> 50.8No clear effect
add Chain-of-Thought instructionsGPT-4o-mini · ResiliBench simulatedfull success rate (%): 50.5 -> 56.1Improved
add Chain-of-Thought instructionsLlama-3.3-70B · ResiliBench simulatedfull success rate (%): 47.8 -> 43.6Hurt
inject systematic flaws into the MDP-optimal workflow prompt7 LLMs average · ResiliBench simulatedfull success rate (%): 62.1 -> 54.3Hurt
inject flaws into workflow promptGemini-2.5-Flash · ResiliBench simulatedfull success rate (%): 60.1 -> 20.0Hurt
inject flaws into workflow promptGPT-4o-mini · ResiliBench simulatedfull success rate (%): 67.7 -> 62.2Hurt
inject flaws into workflow promptGPT-5-mini · ResiliBench simulatedfull success rate (%): 60.7 -> 63.5No clear effect
inject flaws into workflow promptDeepSeek-V3 · ResiliBench simulatedfull success rate (%): 56.8 -> 58.4No clear effect
scale Qwen2.5 from 3B to 7BQwen2.5 with MDP-optimal prompt · ResiliBench simulatedfull success rate (%): 0.5% -> 63.5%Improved
scale Qwen2.5 from 32B to 72BQwen2.5 with MDP-optimal prompt · ResiliBench simulatedfull success rate (%): 65.0% -> 65.0%No clear effect
inject flaws into workflow prompt7 LLMs average · ResiliBench real-world task setfull success rate (%): 46.7 -> 34.8Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy