Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Rejecting Hallucinated State Targets during Planning
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 a minimally-intrusive target feasibility evaluator, trained with an off-policy distributional hitting-time rule plus hindsight relabeling, that rejects hallucinated state targets in planning agents (Skipper, LEAP, Dyna) to reduce delusional behaviors.
Problem
Generative models used as target generators in planning (decision-time and background) hallucinate infeasible state targets, which cause delusional plans and value estimates and safety concerns.
Approach
Attach a learned evaluator that estimates the distribution of hitting times D_pi(s, g) with an off-policy compatible update rule and a C51-style histogram head, so that low probability of reaching a target within tau steps flags it as infeasible; train it with hindsight relabeling augmented by generator-proposed targets (generate) and same-task experienced targets (per-task).
Key results
- Evaluator-assisted Skipper and LEAP show significantly lower feasibility errors and delusional plan frequencies than baselines on SSM and RDS (exact numbers in figures, not in text).
- Evaluator-augmented Dyna shows significantly better performance than baseline Dyna on SSM via rejecting hallucinated next states.
- Results are averaged over 20 seeds with 95% confidence intervals, with improved OOD generalization reported across the four OOD difficulties.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the feasibility evaluator with FEPG relabeling to Skipper | Skipper · SSM OOD evaluation | OOD performance | Helped |
| Add the feasibility evaluator to Skipper | Skipper · SSM | delusional plan frequency | Helped |
| Add the feasibility evaluator with FEPG relabeling | LEAP · RDS OOD evaluation | OOD performance | Helped |
| Add target rejection using the learned feasibility evaluator | 1-step Dyna · SSM | performance | Helped |
| Add target rejection using FEPG | 1-step Dyna · RDS | performance | Helped |
| Use FEG instead of FE for evaluator relabeling (ablation) | Skipper · RDS OOD evaluation | OOD performance | Helped |
| Use FEP instead of FE for evaluator relabeling (ablation) | Skipper · SSM OOD evaluation | OOD performance | Helped |
| Use FEG, FEP, or FEPG instead of FE for evaluator relabeling (ablation) | LEAP · SSM OOD evaluation | OOD performance | Helped |
| Use FEG instead of FE for evaluator relabeling (ablation) | LEAP · SSM OOD evaluation | OOD performance | Helped |
| Increase the share of Generate relabeling (ablation) | Skipper · SSM | E1 feasibility error | Helped |
| Increase the share of PerTask relabeling (ablation) | Skipper · SSM | E2 feasibility error | Helped |
| Train the generator with Future rather than Episodes relabeling (ablation) | HER-trained target generator · SSM | G2 target generation frequency | Helped |
Surprising findings
- Generator hallucinations of temporarily infeasible targets are more frequent when generators are trained with episode-level relabeling than with future relabeling.
Limitations
- Targets considered are state-based; action-like targets are not addressed.
- Generate relabeling adds computational overhead from querying the generator.
- LEAP results on the 12x12 SSM became hard to analyze due to rampant failures, so 8x8 was used.
Experimental setup
Datasets: SSM (custom MiniGrid environment) RDS (custom MiniGrid environment)
Benchmarks: SSM RDS
Metrics: feasibility error delusional behavior frequency OOD generalization success rate
Baselines: Skipper LEAP Dyna DreamerV2
Builds on: Hindsight Experience Replay Dyna C51 Skipper LEAP
Closest papers by meaning
- InEx: Hallucination Mitigation via Introspection and Cross-Modal Multi-Agent Collaboration (AAAI 2026 · similarity 0.86)
- Tracing the Cascade: A Topology-Aware Evaluation Framework for Scientific Agent Hallucinations (NeurIPS 2026 · similarity 0.86)
- BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (NeurIPS 2026 · similarity 0.85)
- Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.85)
- Reasoning Models Hallucinate More: Factuality-Aware Reinforcement Learning for Large Reasoning Models (NeurIPS 2025 · similarity 0.85)
- Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.85)
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.