Skip to content

Adaptive Multi-Agent Response Refinement in Conversational Systems

🕒 Published (v1): 2025-11-11 14:48 UTC · Source: Arxiv · Venue: AAAI 2026 · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MARA (Multi-Agent Refinement with Adaptive agent selection) is a framework that refines LLM conversational responses using three specialized agents—fact-checking, persona alignment, and coherence—orchestrated by a planner agent that dynamically selects both the set and ordering of refinement agents per query. It consistently outperforms single-agent and multi-agent baselines across five conversational datasets, with particularly large gains on tasks requiring factual grounding and persona alignment simultaneously.

Problem

LLMs frequently produce unsatisfactory initial responses in multi-turn conversational settings, especially when responses must jointly satisfy factual accuracy, user persona alignment, and conversational coherence. Single-agent refinement (e.g., Self-Refine) is limited by overconfidence and the inability to holistically address all quality dimensions; prior multi-agent baselines either focus narrowly on factual errors or lack explicitly defined roles for targeted refinement.

Method

MARA deploys four prompt-instantiated LLM agents, all unsupervised (no fine-tuning): 1. Fact-refining agent — corrects factual inaccuracies against provided knowledge. 2. Persona-refining agent — aligns responses with user profile information. 3. Coherence-refining agent — ensures logical consistency with prior conversational turns. 4. Planner agent — given the current query and initial response, generates an ordered sequence s_planner specifying which subset of the three refining agents to invoke and in what order, along with justifications for each decision.

Agents operate sequentially: each agent receives the output of the previous one plus the planner's justification for its role. The backbone is Claude Sonnet 3, with Claude Sonnet 3.5 assigned specifically to the fact-refining agent due to its stronger factual capabilities. Three communication strategies were evaluated—simultaneous (all agents independently → finalizer), static sequential, and the proposed dynamic sequential—with dynamic sequential yielding the best results.

Key Contributions

  • Introduces a three-specialist multi-agent refinement framework covering factuality, persona alignment, and coherence as orthogonal quality axes.
  • Proposes a planner agent that adaptively selects both the composition and ordering of refining agents per query, enabling query-specific refinement without a fixed pipeline.
  • Demonstrates that heterogeneous model assignment across agents (different LLM sizes for different roles) is a concrete practical advantage of multi-agent architectures.
  • Provides ablations over communication strategies (simultaneous vs. static sequential vs. dynamic sequential) and over each agent's individual contribution.

Results

  • On FoCus (persona + knowledge): MARA achieves Overall 74.51 vs. best baseline SPP at 60.47 and No Refine at 56.71.
  • On INSCIT (knowledge-grounded): MARA Overall 72.61 vs. SPP 65.74, MultiDebate 61.70.
  • On PersonaChat (persona): MARA Overall 62.00 vs. Self-Refine 58.41.
  • On Ubuntu Dialogue Corpus: MARA Overall 68.88 vs. best baseline SPP at 51.50.
  • On PRODIGy (role-play, GPT backbone): MARA Overall 63.00 vs. best baseline Self-Refine at 52.13.
  • Ideal planner (brute-forced optimal sequence) achieves higher scores with fewer average agent accesses (3.0) vs. MARA's adaptive planner (4.4), indicating room for planner improvement.
  • Dynamic sequential outperforms all static sequential orderings and the simultaneous strategy on FoCus (Overall 74.38 vs. best static Seq. at ~67-68).
  • MARA generalizes across GPT-4o-mini, LLaMA 3.1 8B, and LLaMA 3.1 70B.
  • Human evaluation on FoCus confirms preference for MARA; G-Eval Spearman ρ with human scores: engagingness 0.58, coherence 0.51, groundedness 0.48, naturalness 0.35.

Limitations

  • Planner agent is unsupervised; a significant gap remains between the current planner and the ideal planner (e.g., FoCus Grd. 0.58 vs. 0.80), suggesting a fine-tuned planner on labeled sequence data could unlock substantial gains.
  • Multi-agent inference increases computational cost; resource efficiency is not fully addressed beyond the observation that a better planner reduces agent calls.
  • G-Eval naturalness shows low human correlation (ρ=0.35), making it an unreliable metric for that dimension.
  • Safety and harm filtering are not integrated; the authors note that adversarial or offensive inputs require additional safeguards not present in MARA.
  • Evaluation is limited to English conversational datasets; cross-lingual or low-resource generalization is untested.

Relevance to Agentic AI / LLM Agents

MARA is a concrete instantiation of the role-specialized multi-agent paradigm applied to a conversational quality control problem, directly demonstrating that agent decomposition by quality dimension outperforms monolithic self-refinement. The planner agent pattern—dynamically routing work to a variable subset of specialized agents based on per-query needs—is a key architectural primitive for agentic orchestration systems more broadly. The finding that heterogeneous model assignment (stronger model for the hardest subtask) amplifies overall system performance has direct implications for cost-optimized multi-agent pipelines. MARA also highlights a persistent challenge in agentic AI: the gap between a learned/prompted planner and an oracle planner, motivating future work on planner training and routing optimization.