Who Verifies the Agents?

Overview

Despite remarkable advances in agents that reason, plan, and operate in open-ended environments, reliably verifying their improvements remains an unsolved challenge. Verification is the ability to determine whether a change to an agent (such as a prompt update, a new tool integration, or a modification to its reasoning process) produces a genuine improvement in capability or reliability.

Verification works in controlled settings (formal math, competitive programming, coding) but remains shallow and noisy for general agentic tasks, causing performance to plateau or silently regress during development. This workshop convenes researchers to tackle verification as a first-class research problem for reliable agent development.

Call for Papers

Topics of Interest

We invite submissions across three core pillars, as well as topics at their intersection:

Pillar 1: Safety and Robustness of Verification

• Robust verifiers that prevent reward hacking and specification gaming
• Adversarial robustness of verifiers and red-teaming of evaluation harnesses
• Alignment-aware verification: ensuring verifiers remain faithful as agents evolve

Pillar 2: Environment-Grounded Verification and Simulators

• Faithful simulators as verification infrastructure for open-ended tasks
• Multi-agent and self-optimizing systems for environment-grounded evaluation
• Measuring agents in production: observability, monitoring, and runtime verification
• Evolutionary and search-based methods for environment-driven agent optimization
• Benchmarks and environment design that stress-test verification methods

Pillar 3: Diverse and Heterogeneous Verifiable Signals

• Composing heterogeneous signals (user experience, cost/latency, calibration, multimodality) into reliable verification metrics
• Beyond scalar rewards: holistic evaluation of agentic behavior
• Human-in-the-loop verification and human–AI collaborative evaluation
• Reflective and self-improving verification (agents that verify other agents)
• Automated agent design, prompt optimization, and scaffold search with verifiable feedback
• Formal verification of agent-generated artifacts, including the use of proof assistants and verification languages (e.g., Dafny, Rocq, and Lean)

Cross-Cutting Topics

• Verification for meta-agents and Agent4Agent systems (agents that design, optimize, or evaluate other agents)
• Self-evolving agents: stable improvement without collapse or reward hacking
• Evaluation of agent-generated designs vs. human-engineered systems
• Scalable oversight and verification for long-horizon, multi-step agent behavior
• Cognitive and neuroscience-inspired verification frameworks

Submission Guidelines

• Format: Papers should be between 4 to 9 pages (excluding references and appendices), using the NeurIPS 2026 template. We also welcome demo papers, which should be no more than 4 pages.
• Dual submission policy: We welcome work that is under review or has been recently published at other venues.
• Review: Reviews will be double blind. Authors of submitted papers may be asked to contribute reviews.
• Presentation: Accepted papers will be presented as posters; select papers will be chosen for oral presentations or lightning talks.
• Non-archival: The workshop is non-archival; accepted papers will be made available on OpenReview but do not constitute formal proceedings.

Submissions and reviewing are handled through OpenReview.