Measuring and Reducing Unusable Replies of LLM Agents
This paper tackles the problem that LLM agents often return replies which a human requester cannot interpret because they rely on private working context unseen by the user. The authors define reply usability as the ability of a requester to understand and act on a reply from only what they have seen ($R$), and measure it with an LLM-based reader model that flags information gaps. Across eight current models, 54% of 480 replies fail, with even the best model failing 30% of the time. They show that appending explicit requirements or rewriting with a secondary model substantially lowers failures, making usability an actionable target for agent design.
The paper makes a strong, timely case that reply usability is a critical and under-measured property of LLM agents, distinct from task success and faithfulness. The authors define failure when a reply 'violates at least one criterion, and the failure rate of a set of replies is the proportion that fail' (Section 3.1), validate their reader model so that it 'agrees with their majority label at Cohen's $\kappa=0.60$ with an accuracy of 82%' (Section 3.2), and show that inexpensive interventions cut failures significantly. However, the reliance on scripted collaborations limits external validity, the heavy use of 2026-dated citations raises questions about the accessibility of the related literature, and the rewriting experiment depends on a single secondary model without cross-model ablation.
The conceptual framing is precise and well-supported: the authors isolate three criteria—interpretability from $R$, correct language, and answering the request—and show that most failures stem from the agent's private context. 'Of the 258 failed replies, 222 contain at least one question that the requester cannot answer from $R$, 73 do not answer the request, and 17 are written in the wrong language, where a reply can meet more than one of these conditions' (Section 5.1). Their theoretical grounding is externally verified: Shaikh et al. (2024) report that 'LLMs generate language with less conversational grounding, instead generating text that appears to simply presume common ground' (abstract), which aligns with the observed gap profile.
The biggest limitation is ecological validity: the collaborations are fully scripted, so the failure rates may not transfer to spontaneous human–agent interactions. The authors acknowledge this directly: 'Our collaborations are scripted, which gives every model the same requests and records the agent's private context that the analysis of gap origins requires' (Limitations).
A second concern is the moderate agreement of the automated judge; while $\kappa=0.60$ is within the human annotator range, it is at the lower bound, and using the reader as ground truth for 480 replies introduces unquantified bias. Additionally, the rewriting intervention uses a single model for all families—'The rewriting model is DeepSeek V4 Pro' (Section 6.3)—so the reported reduction may not generalize to other rewriters or contexts.
The quantitative evidence supports the main claims. The authors show that 'Explicit requirements reduce the failure rate from 0.52 to 0.43, a reduction of 0.098 [0.030, 0.168]' (Section 6.2), and that rewriting cuts the rate to 0.36, which human annotators blindly confirm: 'The share of annotators who fail a reply falls from 0.52 for the default reply to 0.32 for the rewrite, a difference of 0.201 [0.095, 0.321]' (Section 6.3). Comparisons to related work are fair; Table 1 clearly positions usability against task success, faithfulness, hand-offs, and decontextualization, and the authors note that prior work judges reports against the trajectory rather than the requester's view. What is missing is a baseline comparison against simpler automatic metrics such as readability scores, though the authors argue these would not capture information gaps.
Reproducibility is hampered by the absence of publicly released artifacts at the time of writing. While the authors state that 'we will release the collaborations, the reader and its judgments' (Section 1), no repository URL is provided, and the experiments rely on closed API models whose exact checkpoints may not persist. The prompts and settings are documented—'The reader uses Gemini 3 Flash (preview). Appendix A gives the prompts and settings' (Section 3.2)—but hyperparameters such as top-$p$ or system message details for the closed models are not listed. Independent reproduction would require access to the same model versions, the scripted collaboration traces, and the private-context logs, none of which are currently available.
People delegate multi-step work to LLM agents and read only the reply that the agent hands back. Current evaluation scores task completion and the faithfulness of reports, and it leaves open whether the requester can understand and use the reply. We call this property reply usability and measure it with a reader model that takes the requester's position and tests whether the reply can be interpreted from what the requester has seen, is in the requester's language and answers the request. Its agreement with the majority of five annotators lies within the range of the individual annotators. Across eight models from seven families, 54% of 480 replies fail, and the best model still fails on 30%. Failed replies mostly point to material the requester has not seen or use labels that the agent coined during its work. Stating explicit requirements for the reply reduces the failure rate from 0.52 to 0.43 together with errors and omissions, and a secondary model that rewrites the reply reduces it to 0.36, which human annotators confirm in a blind comparison. Reply usability can thus be scored alongside task success and improved with one appended message or one call to a secondary model.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.