Technical Summary: The Yes-Man Syndrome in Embodied Robotic Agents

Problem Statement

Vision-language models (VLMs) are increasingly deployed as high-level planners for embodied robots, translating natural language instructions and visual observations into physical action plans. A critical gap exists in current evaluation frameworks: while abstention (the ability to recognize when a query cannot be reliably resolved) has been studied in text-only Large Language Models (LLMs), existing benchmarks fail to capture the perceptual grounding and physical constraints inherent to robotics.

In embodied settings, "abstention" is defined as withholding execution when an instruction is ill-posed, ungrounded, ambiguous, or physically infeasible given the observed scene and the agent's capabilities. Failure to abstain leads to incorrect task execution, error compounding in multi-step tasks, and degraded interaction. Current systems often default to action even when instructions are ambiguous or impossible, a behavior the authors term the "Yes-Man Syndrome."

Methodology: ROBOABSTENTION Framework

To address this gap, the authors introduce ROBOABSTENTION, a scalable, auditable framework for generating and evaluating abstention instructions grounded in real-world images. The framework operates through a three-phase pipeline designed to ensure programmatic verifiability and reduce hallucination in the dataset generation process:

1. Structured Visual Grounding: A VLM (Anthropic's Claude Sonnet 4.6) extracts a structured JSON representation of the input image. This phase is restricted to perceptual description (objects, attributes, states, spatial locations) using controlled vocabularies. Crucially, the model is prompted to report only what is directly observable and to identify absent or implausible objects, avoiding downstream reasoning at this stage.
2. Deterministic Constraint Derivation: Using rule-based logic on the structured JSON output, the system deterministically derives candidate instances for specific abstention categories. This phase is entirely non-learned, ensuring that the conditions for abstention are mathematically verifiable based on the scene representation.
3. Controlled Instruction Generation: Natural language instructions are generated using category-specific templates. These templates are instantiated only with the candidates derived in Phase 2, ensuring that the generated instructions strictly align with the intended abstention condition (e.g., a "Physical Infeasibility" template is only filled with object-location pairs that violate size constraints).

Taxonomy of Abstention

The framework categorizes abstention into eight distinct types, organized into three groups:

• Reference Grounding:
  • Missing Referent: The object does not exist in the scene.
  • Ambiguous Referent: Multiple objects satisfy the description without distinguishing attributes.
  • Underspecified Intent: Lack of context for deictic references (e.g., "that," "there").
  • Subjective Intent: Requires user preferences unknown to the robot (e.g., "my favorite drink").
• Execution Feasibility:
  • Physical Infeasibility: Violates spatial or physical constraints (e.g., fitting a keyboard in a cup).
  • Missing Capability: Requires sensors or actuators the robot lacks (e.g., smelling).
  • Contradictory Instructions: Imposes mutually incompatible requirements.
• False Premise:
  • False Premise: The instruction assumes a world state contradicted by the scene (e.g., "turn off" a monitor that is already off).

Dataset Construction

The benchmark utilizes 1,250 images sourced from five diverse embodied robotics datasets (DROID, Robo2VLM, RoboVQA, BridgeV2, and EgoThink). The pipeline generates a total of 6,069 instructions, with each image paired with instructions that should elicit abstention across multiple categories. The dataset is designed to be auditable, with every instruction traceable back to the specific visual grounding and constraint derivation that justified its creation.

Experimental Results

The authors evaluated several frontier VLMs, including GPT-5.4, Claude Sonnet 4.6, Gemini Robotics ER 1.6 Preview, and Gemini 2.5 Flash.

• Baseline Performance: All models exhibited significant weaknesses in abstention. The best-performing general model, Gemini 2.5 Flash, abstained on only 39.0% of instructions. The embodied-specific planner, Gemini Robotics ER 1.6 Preview, performed worse, abstaining on just 16.5%.
• Category Disparities: Models performed relatively well on "Missing Referent" cases (41.3% abstention) but failed dramatically on "Ambiguous Referent" tasks (2.6% abstention), suggesting a tendency to guess rather than request clarification.
• Model Scale and Reasoning: Within the GPT-5.4 family, increasing model scale (Nano to Mini to Full) had little effect on overall abstention rates. Surprisingly, increasing the "reasoning" tier (from none to high) reduced abstention performance, indicating that additional deliberation encourages models to construct plausible action plans or workarounds rather than recognizing invalidity.
• Failure Modes: Analysis of non-abstaining responses revealed that models frequently "hallucinate" concrete targets, invent sensory capabilities, or propose search-and-act strategies to bypass uncertainty. The largest failure cluster was "Arbitrary Target Selection," where models resolved ambiguity by picking a target without clarification.

Mitigation Strategies

The study tested whether prompting could improve performance:

• Defensive Prompting: Explicitly instructing the model to abstain under specific conditions.
• In-Context Learning (ICL): Providing few-shot examples of correct abstention behavior.
• Combined Approach: Using both strategies.

The combined approach yielded the highest improvements, raising the abstention rate for Gemini Robotics ER 1.6 Preview to 93.6% and GPT-5.4 Mini to 88.6%. However, the authors note that no single approach fully solved the problem, and significant failure modes persisted even with mitigation.

Significance and Claims

The paper positions abstention as a distinct capability that must be explicitly measured in embodied AI evaluation, separate from task execution success. The authors claim that:

1. Current frontier VLMs, even those with advanced reasoning capabilities, are prone to the "Yes-Man Syndrome," confidently executing instructions that are ambiguous, ungrounded, or physically impossible.
2. Existing benchmarks are insufficient because they lack the visual grounding and physical constraint verification required for robotics.
3. ROBOABSTENTION provides a scalable, auditable, and programmatically verifiable method for generating such benchmarks, moving beyond manual annotation.
4. While prompting strategies can substantially improve performance, the fundamental issue of models defaulting to action under uncertainty remains an open challenge, suggesting that current architectures may lack the inherent mechanisms to recognize their own limitations in physical environments.

The authors conclude that the benchmark should not be interpreted as evidence of readiness for real-world deployment, as it focuses on benign, non-executable instructions and does not cover adversarial misuse or long-horizon interactions. The work aims to support the development of models that can better acknowledge uncertainty and avoid acting under invalid assumptions.