Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment¶

🕒 Published (v1): 2026-04-21 15:05 UTC · Source: Arxiv · Venue: ACL 2026 · link

Why this paper was selected

Dialectical alignment resolves actor-observer asymmetry in multi-agent frameworks [ACL 2026]

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

Role-playing in multi-agent LLM systems induces Actor-Observer Asymmetry (AOA)—a human cognitive bias where actor-role agents externalize blame and observer-role agents internalize it, producing inconsistent failure attribution. The paper introduces ReTAS, a 4B model trained via Thesis-Antithesis-Synthesis (TAS) dialectical chain-of-thought plus GRPO reinforcement to produce perspective-invariant attributions. On the Ambiguous Failure Benchmark (AFB) and two RAG-based attribution datasets, ReTAS reduces AOA flip rates and improves downstream task F1, outperforming models up to 32B.

Problem¶

Multi-agent frameworks assign specialized roles (executor, reviewer) to leverage domain expertise, but role assignment acts as a rigid cognitive prior: the same failure is attributed differently depending on whether the agent is self-reflecting (Actor) or auditing (Observer). Simply prompting agents to "be objective" or adding a dual-view reviewer does not resolve the bias—and can worsen it. No prior work quantified or systematically mitigated this role-induced attribution asymmetry in LLM agents.

Method¶

Ambiguous Failure Benchmark (AFB): 200 interaction traces across 10 domains constructed so the failure cause is genuinely ambiguous (internal reasoning flaw vs. external constraint). Each trace is evaluated twice under Actor and Observer system prompts; inconsistency (flip) measures AOA severity.
Attribution Dataset Construction: Two RAG pipelines over FinQA and Spider produce labeled examples with ground-truth attribution: FalseExt (missing evidence, $E_{gold} \not\subseteq E$), FalseInt (sufficient evidence but wrong answer, $\hat{a} \neq a^*$), or True. This gives an objective reference against which role-induced bias can be measured.
Dialectical Synthesis (TAS): A strong teacher (GPT-5.1) generates two reasoning trajectories per example starting from opposing roles (Defensive Actor, Critical Reviewer). Each trajectory follows: <thesis> (role-induced reaction) → <antithesis> (evidence verification) → <synthesis> (perspective-invariant conclusion + action: Search/Revise/Confirm). Both trajectories must converge to the same attribution label.
Dialectical Alignment (ReTAS): SFT on TAS trajectories initializes format, then GRPO fine-tunes with a composite reward $R = \alpha R_1 + \beta R_2 + \gamma R_3$ ($\alpha=1, \beta=2, \gamma=4$), where $R_1$ rewards correct TAS format, $R_2$ rewards attribution-label accuracy, and $R_3$ rewards downstream answer correctness.

Key Contributions¶

AFB benchmark: Quantifies AOA across 10 domains; shows >20% flip rate for most models, persisting even in GPT-5.1.
TAS reasoning format: Structured three-stage dialectical CoT that decouples role-based reflexes from evidence-grounded conclusions; model-agnostic (consistent gains from 4B–32B).
ReTAS model: Qwen3-4B fine-tuned with SFT + GRPO; achieves state-of-the-art attribution accuracy among open-weight models, surpassing 32B baselines.
Generalization evidence: Cross-domain evaluation on AFB (trained on FinQA-TAS) and dynamic negotiation (Sales Arena) demonstrate TAS transfers beyond the training distribution.

Results¶

AFB: AOA flip rate >20% for most models; Qwen3-4B reaches V-AOA of 29%; GPT-5.1 reduces to ~5% but does not eliminate it.
FinQA-TAS: ReTAS achieves 71.2% attribution accuracy and 5.4% V-AOA vs. best open baseline QwQ-32B (Dual View: 54.9% accuracy, 18.1% Flip, 14.7% V-AOA); F1 = 72.1.
Spider-TAS: ReTAS reaches 61.4% accuracy, 10.2% V-AOA, F1 = 63.5; vs. QwQ-32B Dual View 60.4% accuracy but 34.2% Flip.
Ablation: Removing attribution reward $R_2$ triples V-AOA (5.4 → 16.8); removing GRPO shows SFT alone insufficient (Acc drops from 71.2 → 67.7).
Sales Arena negotiation: Reflection_TAS achieves highest profit ($168 avg, $2.10/turn) vs. Reflection_Dual ($135, $1.69/turn), which performs worse than no reflection at all.

Limitations¶

Evaluation domains are synthetic/constrained (FinQA, Spider); real-world attribution ambiguity may not decompose as cleanly into Internal/External.
ReTAS backbone is 4B (Qwen3-4B); behavior of dialectical alignment at larger scales (e.g., 70B+) is not evaluated.
TAS trajectories are generated by a single teacher model (GPT-5.1), potentially inheriting its own attribution biases.
The AFB dataset contains only 200 traces, limiting statistical power for fine-grained per-domain analysis.
Sales Arena evaluation uses a fixed buyer model (QwQ-32B); generalization to diverse opponent strategies is not assessed.

Relevance to Harnesses / Meta-Harnesses¶

Multi-agent harnesses typically orchestrate specialized role-playing agents (planners, executors, reviewers) in pipelines; this paper directly diagnoses a structural failure mode of that architecture—role-assigned agents cannot self-audit objectively, introducing systematic inconsistency in fault localization. ReTAS's TAS training recipe provides a concrete alignment technique that harness designers could apply to any agent serving a reflection or auditing role within an orchestration pipeline. The finding that naïve dual-view reviewer injection worsens performance without a synthesis mechanism is a direct warning for harness architectures that add critic agents without integrating their outputs. The dialectical alignment paradigm—treating inter-agent disagreement as a structured reasoning primitive rather than a scheduling artifact—is a design principle transferable to harness-level conflict resolution mechanisms.