Source-linked AI summary
Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment
Bobo Li, Rui Wu, Zibo Ji, Meishan Zhang, Hao Fei, Min Zhang, Mong-Li Lee, Wynne Hsu
TL;DR
Role-based self-reflection and auditing induce Actor-Observer Asymmetry, undermining objective consensus. The paper introduces ReTAS, which uses dialectical alignment to reconcile conflicting perspectives, substantially reducing attribution errors without degrading task performance or role-specific capabilities.
Problem
Role assignment induces Actor-Observer Asymmetry, with actors externalizing failures and observers emphasizing internal faults, compromising objective consensus in collaborative agents.
Method
ReTAS trains agents through thesis-antithesis-synthesis reasoning and dialectical alignment to reconcile divergent attributions into perspective-invariant judgments.
Results
ReTAS substantially reduces attribution errors without degrading task performance or role-specific capabilities, while improving attribution bias mitigation and collaborative reliability.
Takeaways & Limitations
Reliable multi-agent systems require alignment of underlying reasoning processes, not merely larger models or surface-level prompt engineering.
Takeaways & Limitations
Evaluation is restricted mainly to FinQA-TAS and Spider-TAS, leaving ReTAS’s effectiveness in long-horizon, creative, and broader real-world settings unresolved.
Abstract
from arXiv · showhide
Large Language Model agents have rapidly evolved from static text generators into dynamic systems capable of executing complex autonomous workflows. To enhance reliability, multi-agent frameworks assigning specialized roles are increasingly adopted to enable self-reflection and mutual auditing. While such role-playing effectively leverages domain expert knowledge, we find it simultaneously induces a human-like cognitive bias known as Actor-Observer Asymmetry (AOA). Specifically, an agent acting as an actor (during self-reflection) tends to attribute failures to external factors, whereas an observer (during mutual auditing) attributes the same errors to internal faults. We quantify this using our new Ambiguous Failure Benchmark, which reveals that simply swapping perspectives triggers the AOA effect in over 20% of cases for most models. To tame this bias, we introduce ReTAS (Reasoning via Thesis-Antithesis-Synthesis), a model trained through dialectical alignment to enforce perspective-invariant reasoning. By integrating dialectical chain-of-thought with Group Relative Policy Optimization, ReTAS guides agents to synthesize conflicting viewpoints into an objective consensus. Experiments demonstrate that ReTAS effectively mitigates attribution inconsistency and significantly improves fault resolution rates in ambiguous scenarios.
1 Introduction
Role-based multi-agent collaboration improves efficiency and solution quality but can induce Actor-Observer Asymmetry, making agents judge identical failures differently by perspective. The paper introduces AFB and ReTAS, which uses dialectical synthesis and reinforcement learning to mitigate attribution bias.
- Motivation: Role-playing assigns specialized expertise and underpins multi-agent frameworks that can outperform monolithic models in efficiency and solution quality.These frameworks mimic human collaboration while distributing tasks across specialized roles.
- Motivation: Assigned roles act as rigid cognitive priors, compromising objectivity during self-correction and peer review.The same failure may receive different judgments depending on whether an agent acts as executor or reviewer.
- Problem formulation: Actors tend to attribute failures to external circumstances, whereas observers attribute them to internal dispositions, defining the targeted AOA bias.The Ambiguous Failure Benchmark tests this bias using scenarios where one failure signature supports contradictory explanations.
- Method motivation: Naïve objectivity instructions or forced opposing perspectives can produce defensive justifications, over-correction, or groundless self-blame because they leave role-induced priors intact.The paper therefore frames dialectical alignment as a response to ambiguity in fault localization.
- Approach and contributions: ReTAS resolves attribution conflicts through Thesis, Antithesis, and Synthesis, then uses consistency-driven reinforcement learning to promote perspective-invariant conclusions.Thesis generates a role-congruent explanation, Antithesis surfaces blind spots, and Synthesis reconciles both views using objective evidence.
- Results and contributions: Experiments indicate that ReTAS significantly mitigates attribution bias and improves task performance, establishing a robust paradigm for agent collaboration.The AFB quantitatively verifies that agent attribution failures mirror human AOA rather than random inconsistency.
2 Related Work
Related work connects the rise of dynamic, role-playing LLM agents to multi-agent task decomposition and domain-specific expertise, while linking their failure-attribution discrepancies to Actor-Observer Asymmetry in social psychology. It further suggests that human-generated training data can transmit such attributional biases to LLMs.
- Role-Playing in LLM Agents: Role-playing multi-agent frameworks use specialized roles to elicit domain expertise and decompose complex tasks.This extends LLMs from static reasoning chains toward dynamic agents with roles such as executor or reviewer.
- Attribution Theory and Cognitive Bias: AOA describes actors attributing failures to situational factors, whereas observers attribute them to dispositional traits.The reported discrepancy in LLM failure attribution mirrors this established pattern in social psychology.
- Attribution Theory and Cognitive Bias: Human-generated training data may transmit attributional biases to LLMs.Prior work is cited as evidence that such biases can be inherited through training on human-generated text.
3 Preliminary Study
The preliminary study introduces the Ambiguous Failure Benchmark (AFB) and a paired Actor–Observer probe to quantify perspective-induced attribution inconsistency. Results show that AOA persists across models, is strongest in smaller models, and remains even in advanced models alongside an attribution imbalance toward internal faults.
- Benchmark design: AFB maximizes attribution ambiguity by withholding deterministic ground truth between internal faults and external factors.The dataset is generated to expose agents’ inherent attribution biases under epistemic uncertainty.
- Benchmark design: AFB spans 10 domains and contains 200 interaction traces: 100 Human-Agent and 100 Agent-Agent traces.The two settings capture ambiguity in user-intent specification versus execution fidelity, and high-level directives versus low-level implementation.
- Evaluation protocol: The evaluation queries each trace twice under identical contexts, changing only the system prompt to induce Actor or Observer roles with forced-choice internal or external attribution.The paired counterfactual probe uses y ∈{Int, Ext}, where Int and Ext denote internal and external causes.
- Findings: AOA persists across all models, with smaller models most prone to externalizing blame as Actors and assigning internal fault as Observers.Flip aggregates V-AOA and R-AOA to measure perspective-induced inconsistency.
- Findings: 29% V-AOA is observed for Qwen3-4B on both benchmarks, while DeepSeek-V3.2 reaches 31% in Agent-Agent scenarios.GPT-5.1 reduces V-AOA to as low as 5%, but the bias remains, indicating scaling alone is insufficient.
- Findings: 94% internal attribution is observed for GPT-5.1 regardless of assigned perspective, indicating an imbalance toward attributing faults to the agent rather than the human user.The passage identifies this pattern as requiring further investigation.
4 Method
The method builds ReTAS through three stages: generating verifiable failure-attribution data, synthesizing dialectical reasoning trajectories, and aligning the model on those trajectories. Its design trains agents to ground attribution and corrective actions in evidence rather than role-induced reactions.
- Overview: The three-stage pipeline generates diagnostic attribution data, converts cases into dialectical trajectories, and trains ReTAS through dialectical alignment.The stages are Attribution Data Generation, Dialectical Synthesis, and Dialectical Alignment.
- Attribution Data Generation: Retrieval-augmented reasoning separates evidence retrieval from answer generation, enabling failures to be classified by verifiable causes.Missing necessary evidence is labeled FalseExt, sufficient evidence with an incorrect answer FalseInt, and sufficient evidence with a correct answer True.
- Dialectical Synthesis: TAS records a role-induced Thesis, an evidence-checking Antithesis, and a conflict-resolving Synthesis that outputs an attribution label and corrective action.A strong teacher model, GPT-5.1, generates trajectories that include potentially incorrect initial reactions and subsequent verification.
- Dialectical Synthesis: Opposing Defensive Actor and Critical Reviewer trajectories are required to converge on the same evidence-grounded attribution.This contrasts role-induced biases and reinforces that final attribution should depend on evidence rather than the assigned role.
- Dialectical Alignment: ReTAS is trained in two phases: supervised fine-tuning teaches the TAS format, followed by GRPO reinforcement learning that turns dialectical reasoning into a behavioral habit.The reinforcement phase scores rollouts for TAS format, attribution-label correctness, and answer correctness.
5 Experiments
Experiments show that ReTAS reduces role-induced attribution bias and improves downstream reliability across ambiguous-failure, cross-model, evidence-complexity, human-agent, and negotiation settings. Ablations indicate that attribution rewards, answer-correctness rewards, and GRPO jointly support its dialectical policy.
- Main results: ReTAS consistently outperforms the evaluated baselines on FinQA-TAS and Spider-TAS, achieving state-of-the-art attribution accuracy and flip scores among openweights models with 4B parameters.The comparisons include standard prompting, single-view reflection, and dual-view reflection baselines.
- Main results: Dual View reflection can underperform Single View reflection, whereas ReTAS reduces V-AOA and bridges conflicting perspectives.The results indicate that assigning opposing roles alone does not resolve role-induced cognitive bias.
- Ablations: 5.4 →16.8 V-AOA follows from removing the attribution reward, while removing answer correctness impairs F1 performance and removing GRPO widens the gap from the full model.These ablations support multi-objective optimization and identify GRPO as critical for learning the dialectical policy.
- Model and prompting generalization: 22.7%/22.2% V-AOA under Dual View falls to 14.1%/15.6% with zero-shot TAS on FinQA-TAS/Spider-TAS, while trained ReTAS achieves the highest attribution accuracy and lowest bias.TAS improvements generalize across models ranging from 4B to 32B and consistently outperform standard Dual View reflection.
- Evidence complexity: TAS-based methods outperform standard Dual View models when reasoning from 1–2 evidence pieces, while zero-shot QwQ-32B degrades sharply as evidence complexity reaches 3 and 4+ pieces.The evidence-complexity results suggest structured dialectical reasoning is especially useful when context is concise.
- Cross-setting generalization: ReTAS reduces Agent-Agent attribution bias, assigns responsibility from evidence rather than user preference in Human-Agent settings, and reaches V-AOA of 11 on an Agent-Agent benchmark.In negotiation, Reflection_Dual reduces total profit to $135, whereas Reflection_TAS achieves the highest profit with fewer negotiation turns and more adaptive strategy formation.
6 Conclusion
The paper identifies Actor-Observer Asymmetry as a systematic bias in role-playing language agents and proposes ReTAS, whose dialectical alignment reduces attribution errors while preserving performance and role capabilities.
- Conclusion: Actor-Observer Asymmetry is a systematic cognitive bias inherent to role-playing language agents.The paper frames this bias as a central challenge for multi-agent systems.
- Conclusion: Functional specialization creates a trade-off with objective consensus: executors externalize blame, whereas auditors overemphasize internal reasoning faults.The asymmetry arises from agents occupying different functional roles.
- Conclusion: ReTAS applies dialectical alignment to reconcile reasoning across divergent perspectives.Its approach targets attribution inconsistency by structuring the reasoning process.
- Conclusion: Structured dialectical reasoning substantially reduces attribution errors without degrading task performance or role-specific capabilities.The result supports dialectical alignment as a way to mitigate bias while retaining agent functionality.
- Conclusion: Increasing model scale alone is insufficient to resolve social-cognitive biases; reliable multi-agent systems require alignment of the underlying reasoning process.The conclusion motivates a shift from surface-level prompt engineering toward principled cognitive alignment.
Limitations
The study’s diagnostic testbed is limited to FinQA-TAS and Spider-TAS, leaving ReTAS’s effectiveness in complex, subjective, and broader real-world settings for future exploration.
- Diagnostic testbed: The evaluation is restricted to FinQA-TAS and Spider-TAS, simplifying the open-ended decision spaces of fully autonomous agents.This restriction supports internal validity but narrows the diagnostic scope.
- Unexplored scenarios: ReTAS’s efficacy in long-horizon planning and creative generation remains unexplored because objective fault attribution is inherently subjective in these scenarios.These settings differ from the structurally isolated benchmark tasks used in the study.
- Negotiation scope: The Sales Arena study provides an initial probe of multi-turn negotiation, while broader real-world negotiation settings remain future work.The existing study does not establish performance across the wider range of real-world negotiation environments.
Ethical Considerations
The study synthesizes human-agent conflict scenarios that may generate toxic or discriminatory content, while human verification audits synthetic traces to keep simulated behaviors within safe operational boundaries.
- Risk of harmful content: The Ambiguous Failure Benchmark synthesizes human-agent conflict scenarios, raising concerns about toxic or discriminatory content generation.The stated objective is to simulate cognitive causal ambiguity rather than semantic toxicity.
- Human oversight: A human-in-the-loop protocol audits a statistically significant subset of synthetic traces.This verification checks that simulated defensive behaviors remain within safe operational boundaries.
- Safety boundaries: The safety audit aims to prevent harmful social stereotypes and offensive language from propagating.These safeguards define the operational boundaries for the simulated behaviors.
A Supplementary Experiment Details
The supplementary experiments describe the training setup, reward-weight robustness, and dataset construction for ReTAS. All three reward components are necessary, while reward-ratio variation has only a modest effect on performance.
- Hardware and Training Time: Experiments used dual NVIDIA H200 GPUs on Ubuntu; SFT took approximately 15 minutes per epoch and GRPO required 9 hours at sequence length 2,048.The setup prioritized algorithmic robustness over parameter engineering and adopted the reward weight ratio without extensive tuning.
- Reward Weight Sensitivity: Reward-ratio variation had a modest performance impact, but ablating any component caused notable degradation, confirming all three components are indispensable.The analysis varied non-zero ratios and separately removed individual reward components.
- ReTAS Training Dataset Statistics: ReTAS training samples from FinQA and Spider were stratified into External, Internal, and Correct categories to balance failure-mode coverage.FinQA used train for training, dev for validation and checkpoint selection, and test for main results; Spider used its publicly released train and dev splits.
B Prompts and Examples
The appendix specifies prompts for generating ambiguous Human-Agent failures and for simulating actor-observer reasoning across fault types. It also defines the TAS reasoning format, structured outputs, and a negotiation example showing improved strategy after dialectical reflection.
- B.1 AOA Dataset Generation Prompt: The generator creates realistic grey-area Human-Agent interactions whose failures may be attributed to either the human or the agent.It follows the “Literal vs. Pragmatic Gap” construction logic.
- B.1 AOA Dataset Generation Prompt: The dataset statistics distinguish External, Internal, and Correct categories across the ReTAS training datasets.
- B.1 AOA Dataset Generation Prompt: The generated scenarios include coding cases where ambiguous interpretations, such as “clean up,” make fault attribution debatable.
- B.2 System Prompt Designs by Fault Type: The system prompts implement Type 1 External Fault with the Reviewer as observer and Executor as actor, while Types 2 and 3 retain the TAS structure with different attribution targets.
- OUTPUT FORMAT (JSON): The JSON output records the scenario, interaction history, actor and observer prompts, neutral question, responsibility label, and search-query candidates.
- OUTPUT FORMAT (JSON): The observer prompt requires concise reasoning, hidden-ground-truth verification, and constructive identification of the exact missing context to retrieve.
- OUTPUT FORMAT (JSON): The prescribed TAS sequence begins with direct criticism, pivots through “Wait...” or equivalent context checking, and concludes with external attribution plus specific missing information.
- OUTPUT FORMAT (JSON): A complete Sales Arena example shows a first-round deadlock followed by dual TAS reflection and an improved strategy in Round 2.
B.3 Sales Arena: Multi-Round Negotiation Experiment
The Sales Arena evaluates reflection mechanisms in dynamic, multi-round commercial negotiations between specialized seller agents and a tough-n user-controlled buyer. It compares four reflection settings using profitability, efficiency, and deal-completion metrics.
- Experimental Setup: The Sales Arena simulates commercial negotiations in a multi-agent framework to validate reflection mechanisms during dynamic, multi-round interactions.The framework includes an Actor conducting negotiations and a Reviewer analyzing history to adjust strategy.
- Experimental Setup: The simulation covers 4 distinct items, with a buyer budget of $260, seller unit cost of $50, and target price of $65+.The Seller Team consists of an Actor and Reviewer, while an independent LLM controls the buyer as a tough negotiator.
- Experimental Setup: Offers between $55 and $65 trigger aggressive bargaining, offers above $75 cause immediate rejection, and each item allows at most 8 negotiation turns.These conditions define the negotiation environment and turn constraint.
- Comparative Reflection Methods: The comparison includes NONE, Reflection_SOLO, Reflection_Dual, and Reflection_TAS (Ours), with TAS using Thesis, Antithesis, and Synthesis.NONE has no reflection; SOLO uses Actor self-reflection; Dual uses debate between Actor and Reviewer; TAS applies a Fichtean dialectic framework.
- Evaluation Metrics: Performance is assessed by Total Profit, Avg Profit/Product, Avg Turns, and Success Rate.These metrics capture cumulative profit, average margin per item, efficiency, and deals concluded within the turn limit.