Evidence-inference reconstruction improves lLM reasoning accuracy with fewer retries
Evidence-Inference Reconstruction: When The Evidence Is Recalled But The Reasoning Goes Wrong
Artificial Intelligence
Summary
Some AI systems need to check their intermediate steps carefully before giving a final answer, which can be slow and costly because they repeat these steps many times. The authors propose a method called Evidence-Inference Reconstruction (EIR) that gathers all the necessary evidence first and then makes one final call to get the correct answer, even if some earlier reasoning was flawed. By doing this, EIR improves how often the AI gets the right answer and reduces how many times it has to run processes. The authors tested EIR with popular AI models and showed it works better and more efficiently than previous retry methods.
What this means in practice
- •For ai system engineers: Build AI reasoning agents that retrieve and consolidate evidence before a single final answer call, reducing computation and improving accuracy.
- •For content retrieval teams: Enhance multi-step question answering systems by separating evidence gathering from final inference, achieving more reliable responses with fewer retries.
Authors
Megan Diehl, Ser-Nam Lim
Abstract
Modern multi-hop LLM agents are equipped with built-in mechanisms to detect errors in intermediate reasoning steps. Such errors trigger corrective actions from these agents, which mostly follow the paradigm of retrying the steps or the reasoning trajectories. Not only are these retries expensive, we present in this paper that they are also potentially unnecessary. To this end, we introduce Evidence-Inference Reconstruction (EIR), which uses structured state to guide one retrieval trajectory, accumulating source evidence in the process. We show that as long as the relevant evidence has been collected, EIR is capable of generating the correct answer in a single final model call even if erroneous evidence has been mixed in due to incorrect intermediate reasoning steps. In one evaluation, using Haiku 4.5 and GPT-4.1 Mini, we evaluate EIR on matched 1,000-question subsets of HotpotQA, 2WikiMultiHopQA, and MuSiQue, showing that EIR improves Answer F1, the overlap between the model's and the correct answer, over the baseline by 8.3--32.8 points, Agentic SSR by 10.6--29.1 points, and Reflexion by 1.1--15.9 points. Additionally, we show that EIR averages 4.85 total model calls per question, compared with 35.29 for Agentic SSR and 12.41 for Reflexion. Together, these results corroborate EIR's central premise: separating evidence retrieval from the final answer model call can improve answer accuracy while utilizing substantially less computation.