acceptodds
Withdrawn from ICLR 2027

Refuse the Harm, Complete the Task via Graph-of-Rules Planning for LLM-Enabled Robots

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large Language Models (LLMs) are increasingly used in embodied AI for robotic manipulation and planning, but their deployment in physical systems exposes vulnerabilities such as jailbreak attacks that can induce unsafe behavior. Existing inference-time defenses are mainly evaluated on short instructions containing a single hazard, whereas realistic tasks may interleave safe and unsafe sub-tasks, conceal hazardous steps, or combine individually safe actions into unsafe outcomes. Such compound instructions make it harder for a single safety decision to distinguish hazardous sub-tasks from legitimate ones, increasing the risk of either overlooking the hazard or rejecting the entire instruction. Thus, we introduce safe task completion, a task that requires an LLM-enabled robot to refuse the harmful part of an instruction while completing its legitimate part. To address this problem, we propose RuleGraph, a multi-agent Graph-of-Rules planning framework that distributes safety enforcement across multiple reasoning and verification stages. Five cooperating agents are used to ground general safety principles into world-specific rules, assess the instruction's sub-tasks in context, and derive a safe goal from the legitimate ones. Planning toward this goal is then organized as a graph of rule-specific nodes, where higher-level nodes inherit the union of their parents' rules, and candidate plans are progressively verified and aggregated until a single plan satisfies every relevant rule. In this way, RuleGraph reduces reliance on a single centralized safety decision. We evaluate RuleGraph on pure-attack and compound instructions across two benchmarks in two different environments using Attack Success Rate (ASR) and Task Success Rate (TSR). RuleGraph reduces ASR to 0% on both benchmarks while achieving the highest TSR among all methods, reaching 98.57% on the RoboGuard benchmark and 84.54% on our long-horizon compound instructions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, AI Safety, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_4qwVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces safe task completion for LLM-enabled robots: rather than accepting or rejecting a mixed-intent instruction as a whole, the robot should refuse harmful sub-tasks while completing legitimate ones. The proposed RuleGraph framework uses five logical agents to ground and decompose the instruction, instantiate world-specific safety rules, classify sub-tasks in full-instruction context, synthesize a safe goal, and plan toward that goal. Planning is organized as a hierarchical graph in which each leaf node handles one rule and higher-level nodes progressively aggregate their parents’ rule sets. Candidate plans undergo feasibility checks, rule-specific generate-judge-refine loops, and a final gate for goal completion and residual malicious intent. The method is evaluated on 70 pure-attack instructions from the RoboGuard benchmark and a new Compound set of 97 mostly mixed-intent instructions. With GPT-4o, RuleGraph reports 0% ASR and 98.57% TSR on RoboGuard, and 0% ASR and 84.54% TSR on Compound, outperforming an undefended planner, a single-plan judge, RoboGuard, and SafePlan. Additional validation includes an adaptive white-box attack, a Gemini-based experiment on RoboGuard, 40 purely safe instructions, and a single-node graph ablation.

Critical strengths

The paper identifies an important limitation of binary guardrails and formulates a practically meaningful alternative. RuleGraph provides a coherent defense-in-depth architecture and shows strong results on the setting it targets.

Critical weaknesses

The central Compound evaluation is small and insufficiently transparent. The paper’s main contribution—safe completion of mixed-intent instructions—is evaluated primarily on an author-constructed dataset of only 97 instructions. The paper provides insufficient detail about dataset generation, category coverage, annotation and adjudication procedures, inter-annotator agreement, or whether the method and its prompts were refined after inspecting benchmark failures.

RuleGraph appears to use substantially more inference-time computation than the baselines, but the comparison is not compute-matched. RuleGraph creates a separate generate-judge-refine loop for every relevant rule, allows up to seven iterations per node, hierarchically aggregates the resulting nodes, and adds several instruction, goal-synthesis, verification, and final-checking stages. It may also revise the goal and rebuild the graph. However, the paper reports no LLM-call counts, token usage, latency, or monetary cost, and does not compare against baselines given a similar inference budget. Consequently, the reported gains cannot be confidently attributed to the graph-of-rules structure rather than simply to substantially greater test-time computation and repeated self-verification. The omission also leaves the practicality of the approach for latency-sensitive robotic systems unclear.

Questions for authors
  1. Please provide full details of the Compound dataset construction and evaluation procedure, including instruction sources and category distribution, annotation and adjudication protocols, annotator agreement, who labeled generated plans as ASR/TSR/neither, and whether any method prompts were modified after inspecting benchmark examples or failures.
  2. Please report the number of LLM calls, input/output tokens, latency, and estimated cost for each method. Do RuleGraph’s advantages remain when the single-plan baseline or other baselines are given a comparable inference-time compute budget?
Reviewer_YLCUOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In the paper, the authors propose a new method called RuleGraph, which is an LLM-based framework for improving the safety of robotic task planning. The main problem the authors are trying to solve is that existing safety methods either reject an entire instruction or accept it, even when the instruction is both safe and unsafe tasks, which can be very harmful. The authors say that a robotic system should instead be able to refuse only the harmful parts but at the same time still completing the legitimate ones. In order to solve this, the authors propose a multi-agent framework that has five LLMs that are responsible for different instruction decomposition like safety-rule generation, goal selection, planning, and verification. The experiments evaluate their method against RoboGuard and SafePlan on two benchmarks and report a 0% attack success rate, along with substantially higher task success rates than the baselines.

Critical strengths
  • The paper is well written and easy to follow.
Critical weaknesses
  • I think the novelty of the paper is very limited, and since the framework is also very simple like five llm calls in a pipeline. I feel that the use of terminiology is also very wrong, since the method is an llm pipeline with fixed llm calls rather than an llm agent network where there is some decision making on the fly on which llm is to be called, etc. Also the use of the term graph-of-rules does not at all seem relevant to me, since there is requirement of graph structures, like dependencies, etc.

  • I very strongly belief that the paper is not relevant for todays models, since even in the experiments the models used are GPT4o, these problems of model not understanding the safety aspect might not be relevant for the recent reasoning models, since the models are atleast a couple of orders of magnitude smarter and they would know what are unsafe actions.

  • Also the main finding of the paper that distributing task load like safety and planning, etc to different agent increase the success rate and reduces failures is already well known and published in papers not cited by the authors [1,2,3].

  • Experiments seem in-conclusive to me without using recent models where we can verify if the problem is even relevant or not.

[1] Khan et al., SAFER: Safety Aware Task Planning via Large Language Models in Robotics. IROS 2025. [2] Li et al., Safe Planner: Empowering Safety Awareness in Large Pre-Trained Models for Robot Task Planning. AAAI 2025 [3] Son et al., SAFEL: Subtle Risks, Critical Failures. EMNLP 2025.

Questions for authors
  • Experiments use GPT-4o. Do the safety failures that motivate the framework persist with recent reasoning models such as o3, GPT-5, Claude, or DeepSeek-R1? Please report results with at least one such model, both as a single-agent baseline and inside your framework. If a strong reasoning model alone already avoids most unsafe actions, what does the multi-agent decomposition still contribute?
Reviewer_pPddOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RuleGraph, a framework for rejecting harmful parts of user instructions while completing their legitimate parts. The system grounds general safety principles in a symbolic world, identifies legitimate subtasks, and derives a safe goal. Planning is done by using a hierarchy of nodes that generate and refine candidate plans combined with safety rules. LLM judges check rule compliance and task completion, and action preconditions are verified deterministically. Experiments on the RoboGuard benchmark and a curated dataset evaluate generated plans and refusals, reporting lower attack success rates and higher task success rates than the compared baselines.

Critical strengths
  • The paper tackles a timely and relevant topic: safety in robotic task planning with LLMs. Evaluating these systems beyond task completion is important. Preserving legitimate parts of a request while rejecting harmful ones is a useful formulation of the problem.
  • The comparison with three defended baselines and an undefended planner provides useful empirical evidence.
Critical weaknesses
  • Unclear research motivation relative to existing LLM–formal-method systems. The paper identifies errors in LLM-generated safety specifications and difficulties handling mixed-intent instructions. However, these limitations primarily motivate improving specification generation and instruction interpretation; they do not explain why the resulting system should rely fully on LLM judgments for safety. The conclusion then proposes incorporating formal verification to get stronger guarantees, pointing back toward an architecture already established in prior work. This makes the research trajectory difficult to understand: what capability does RuleGraph add that could not be developed within an existing hybrid framework? Instruction decomposition and natural-language-to-formal-specification translation are not themselves new, so the authors need to identify the specific gap addressed by their rule hierarchy.

    This concern also applies to planning. The system uses symbolic world states and deterministic action-precondition checks. Although the paper does not provide a complete formal action domain, these assumptions suggest a natural role for symbolic planning (either PDDL or STRIPS). The reported improvements establish the performance of the proposed method on the selected benchmarks, but do not sufficiently justify its architecture in relative to extending an existing hybrid system.

  • Unclear formulation of tasks, goals, and planning. The example inputs already specify plan-ish action sequences, which the system converts into a natural-language goal before generating another plan. Although the goal-agent prompt provides qualitative guidance, the resulting goal remains a natural-language description without explicitly defined satisfaction conditions. Thus, even before removing unsafe content, the planning problem and its correctness criteria are insufficiently specified. Safety-driven revision compounds this issue. The prompt permits the goal to diverge from the request, allows reordering, and permits omitting legitimate outcomes judged unrealizable. The paper therefore needs to define how the original task is interpreted and transformed into a safe task, (such as Goriginal→GsafeG_{\mathrm{original}} \rightarrow G_{\mathrm{safe}}), including which actions, ordering constraints, and outcomes must be preserved. This also affects evaluation: although TSR requires completion of all legitimate subtasks, it is unclear how those requirements are established when goal is diverged.
  • Limited validation of the robotics claims. The experiments were run on semantic graphs, with no physical robot experiments or simulation. This abstraction leaves perception, motion execution, and the correspondence between symbolic states and physical conditions untested. It also limits what the evaluation can establish about safety-rule coverage. For example, hazards involving a knife’s orientation relative to a person or its placement within a child’s reach. These details might be only observed when being executed in an embodied environment.
  • Insufficient detail about the safety specifications. The paper describes generating world-specific rules from predefined general principles (P), but it’s unclear what safety knowledge the system receives. The fixed general principles PP and example rules RwR_w should be provided too for better understanding.

Other comments:

  • Formal methods for LLM-agent safety has been studied before Roboguard, and the paper should credit papers including:
    • Plug in the Safety Chip: Enforcing Constraints for LLM-driven Robot Agents.
    • Ensuring Safety in LLM-Driven Robotics: A Cross-Layer Sequence Supervision Mechanism.
Questions for authors
  • For the oven example (G(!switchoff(oven))), could the authors provide the complete prompts and inputs used to generate RoboGuard’s LTL formula and RuleGraph’s safety rules? Were both methods given equivalent safety principles, world information, and action descriptions?
  • Could the authors provide the action model, including skill preconditions and effects, complete prompts for all agents, the fixed principles (P), and representative input instructions with their generated and selected safety rules? Please distinguish human-specified requirements from constraints inferred by the LLM.
  • How is TSR evaluated when the system can revise its goal or omit outcomes judged unrealizable? Who independently determines the legitimate subtasks and their required outcomes or ordering constraints? Does omitting a legitimate subtask count as failure even when the generated safe goal is achieved?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.