Research map / Multi-agent systems and games / LLM agents

How can we assess human-agent interactions? Case studies in software agent design

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PULSE, a prediction-powered framework that collects in-the-wild user star ratings for coding-agent sessions, trains a feature-based model to infill missing ratings, and uses PPI to compare agent designs with tighter CIs, deployed on OpenHands with 15k users.

Problem

Benchmarks assume full automation and do not capture human-agent collaboration; sparse explicit user feedback makes rigorous, efficient human-centric evaluation of agent designs hard.

Approach

Combine per-work-segment user ratings with an ML model that predicts satisfaction from interpretable trajectory features, and use prediction-powered inference to get valid, tighter confidence intervals for A/B comparisons of agent designs.

Key results

• Random forest satisfaction predictor reaches correlation 0.29 and MAE 1.02, outperforming LLM-as-a-judge baselines (correlation 0.14-0.23).
• PULSE reduces 95% CI widths by about 39.5% on average versus naive A/B tests across four experiments.
• Claude-4-sonnet is rated 5.86% higher than claude-3.7-sonnet, while gpt-5 is rated 7.83% lower than claude-4-sonnet in user satisfaction.

Surprising findings

• Human ratings and benchmark differences are negatively correlated for Claude 4 vs GPT-5 (rho = -0.18 in table, -0.11 in text), despite gpt-5 winning 6 of 7 benchmarks.
• Scaffolding changes such as planning and memory settings have much smaller effects on satisfaction than the choice of LLM backbone.

Limitations

• Single agent platform (OpenHands) limits generality of design findings.
• Label noise is not explicitly modeled; only about 5% of sessions receive ratings.
• Raw code contexts cannot be released due to privacy constraints.

Experimental setup

Datasets: in-the-wild OpenHands user sessions with 5-star ratings SWE-Bench Verified subset (for memory parameter simulation)

Benchmarks: SWT-Bench The Agent Company SWE-Bench

Metrics: MSE MAE correlation confidence interval width effect size (Delta)

Baselines: LLM-as-a-judge with o3 LLM-as-a-judge with gemini-2.5-pro LLM-as-a-judge with claude-4 naive A/B test difference in means

Builds on: prediction-powered inference OpenHands A/B testing

Closest papers by meaning

• Talk, Evaluate, Diagnose: User-aware Agent Evaluation with Automated Error Analysis (ICLR 2026 · similarity 0.88)
• General Agent Evaluation (NeurIPS 2026 · similarity 0.87)
• ProSoftArena: Benchmarking Hierarchical Capabilities of Multi-modal Agents in Professional Software Environments (CVPR 2026 · similarity 0.86)
• AppWorld-UL: Benchmarking Diverse Agent-User Interactions for Tool-Use (ICML 2026 · similarity 0.86)
• Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation (ICLR 2026 · similarity 0.86)
• REAL: Benchmarking Autonomous Agents on Deterministic Simulations of Real Websites (NeurIPS 2025 · similarity 0.86)