VestaBench: An Embodied Benchmark for Safe Long-Horizon Planning Under Multi-Constraint and Adversarial Settings (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Existing safety benchmarks do not represent a diverse range of multi-constraint tasks that require long-horizon planning with a focus on safety. |
| Approach: | They propose a benchmark to assess the safety of embodied AI agents under multiple constraints. |
| Outcome: | The proposed benchmarks show that LLMs perform poorly against their tasks . they also suffer significantly compromised safety outcomes . |
Similar Papers
DeepPlanning: Benchmarking Long-Horizon Agentic Planning with Verifiable Constraints (2026.acl-long)
Copied to clipboard
Yinger Zhang, Shutong Jiang, Renhao Li, Jianhong Tu, Yang Su, Lianghao Deng, Xudong Guo, ChenXu Lv, Junyang Lin
| Challenge: | Existing LLM planning benchmarks emphasize local, step-level reasoning rather than global constrained optimization. |
| Approach: | They propose a benchmark for practical long-horizon agent planning that uses local constrained reasoning and global constrained optimization. |
| Outcome: | The proposed benchmarks show that even frontier agentic LLMs struggle with these problems. |
SafetyALFRED: Evaluating Safety-Conscious Planning of Vision Language Models (2026.findings-acl)
Copied to clipboard
Josue Torres-Fonseca, Naihao Deng, Yinpei Dai, Shane Storks, Yichi Zhang, Rada Mihalcea, Casey Kennington, Joyce Chai
| Challenge: | Existing safety evaluations focus on hazard recognition through disembodied question answering (QA) settings, but lack a critical gap in evaluating an agent. |
| Approach: | They evaluate multimodal large language models with six categories of kitchen hazards . they propose a safety-based approach that prioritizes multi-step corrective actions . |
| Outcome: | The proposed model can recognize hazards in QA settings, but average mitigation success rates are low . the proposed model is based on the embodied agent benchmark ALFRED . |
MADRA: Multi-Agent Debate for Risk-Aware Embodied Planning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing safety alignment methods, such as RLHF, fall into a Safety-Utility Trade-off, resulting in severe over-rejection of benign household instructions. |
| Approach: | They propose a meta-cognitive Critical Agent that evaluates peer debates using a structured argumentation framework derived from the Toulmin Model. |
| Outcome: | The proposed architecture outperforms existing systems in the SafeAware-VH benchmark. |
TAMAS: Benchmarking Adversarial Risks in Multi-Agent LLM Systems (2026.acl-long)
Copied to clipboard
| Challenge: | Existing benchmarks and datasets focus on single-agent settings, failing to capture the unique vulnerabilities of multi-agend LLM dynamics and co-ordination. |
| Approach: | They propose a benchmark to evaluate the robustness and safety of multi-agent LLM systems. |
| Outcome: | The proposed benchmark evaluates the robustness and safety of multi-agent LLM systems. |
LongCLI-Bench: A Preliminary Benchmark and Study for Long-horizon Agentic Programming in Command-Line Interfaces (2026.findings-acl)
Copied to clipboard
Yukang Feng, Jianwen Sun, Zelai Yang, Jiaxin Ai, Chuanhao Li, Zizhen Li, Fanrui Zhang, Kang He, Rui Ma, Jifan Lin, Jie Sun, Yang Xiao, Sizhuo Zhou, Wenxiao Wu, Yiming Liu, Pengfei Liu, Shenglin Zhang, Kaipeng Zhang
| Challenge: | Existing benchmarks for agentic programming in long-horizon command-line interface tasks are limited by short task horizons, data contamination from GitHub scraping, and a lack of fine-grained evaluation metrics. |
| Approach: | They propose a benchmark to evaluate agentic capabilities across long-horizon command-line interface tasks. |
| Outcome: | The proposed benchmarks cover four engineering categories: from scratch, feature addition, bug fixing, and refactoring. |
Safety Is Not Universal: The Selective Safety Trap in LLM Alignment (2026.findings-acl)
Copied to clipboard
Iago Alves Brito, Walcy Rios, Julia Soares Dollis, Diogo Fernandes Costa Silva, Arlindo Rodrigues Galvão Filho
| Challenge: | Existing safety evaluations of large language models aggregate harms under generic categories such as "Identity Hate" a bilingual benchmark identifies a selective safety trap, where defense rates vary by up to 42% within the same model solely based on the target group. |
| Approach: | They propose a bilingual adversarial benchmark to audit selective safety in large language models . defense rates vary by up to 42% within the same model solely based on target group . |
| Outcome: | The proposed benchmark identifies a selective safety trap in large language models . defense rates vary by up to 42% within the same model solely based on the target group. |
Subtle Risks, Critical Failures: A Framework for Diagnosing Physical Safety of LLMs for Embodied Decision Making (2025.emnlp-main)
Copied to clipboard
Yejin Son, Minseo Kim, Sungwoong Kim, Seungju Han, Jian Kim, Dongju Jang, Youngjae Yu, Chan Young Park
| Challenge: | Existing safety evaluations rely on coarse success rates and domain-specific setups, making it difficult to diagnose why and where these models fail. |
| Approach: | They propose a framework for systematically evaluating the physical safety of LLMs in embodied decision making. |
| Outcome: | The proposed framework assesses the physical safety of LLMs in embodied decision making. |
Beyond Reactive Safety: Risk-Aware LLM Alignment via Long-Horizon Simulation (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing alignment methods focus on reactive feedback, where immediate human perception is leveraged to judge sampled model responses as preference data for post-training. |
| Approach: | They propose a proof-of-concept framework that projects how model-generated advice could propagate through societal systems on a macroscopic scale over time, enabling more robust alignment. |
| Outcome: | The proposed framework achieves 20% improvement on existing safety benchmarks and an average win rate exceeding 70% against strong baselines. |
LCO: LLM-based Constraint Optimization for Safer Agentic LLMs in Real-world Tasks (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing defense methods are insufficient to address in-context reward hacking (ICRH), where LLMs iteratively optimize their behavior to maximize proxy objectives, resulting in harmful side effects. |
| Approach: | They propose a framework that reduces in-context reward hacking (ICRH) through repeated interactions with the environment. |
| Outcome: | The proposed framework reduces ICRH without model fine-tuning while maintaining task performance. |
OmniCompliance-100K: A Multi-Domain, Rule-Grounded, Real-World Safety Compliance Dataset (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing LLM safety datasets rely on ad-hoc taxonomies and lack rule-grounded, real-world cases. |
| Approach: | They construct a rule-grounded, real-world case dataset OmniCompliance-100K from a compliance perspective using a powerful web-searching agent. |
| Outcome: | The proposed dataset spans 74 regulations and policies across a wide range of domains including security and privacy regulations, content safety and user data privacy policies, financial security requirements, medical device risk management standards, educational integrity guidelines, and protections of fundamental human rights. |