Multi-Agent MLLM Judge Framework

• Multi-Agent MLLM-as-a-Judge is a system where multiple specialized LLM agents collaboratively decompose evaluation tasks into coordinated judgments using defined protocols and aggregation functions.
• Architectural patterns vary from sequential inspector chains to multi-round debates, enabling precise tool planning, multi-dimensional scoring, and adaptive prompt optimization.
• Protocol design and aggregation strategies, including weighted averages and consensus checks, enhance reliability, mitigate biases, and drive efficiency in multi-turn evaluations.

Multi-Agent MLLM-as-a-Judge (MAM) denotes an evaluation paradigm in which judgment is produced by a coordinated set of LLM- or MLLM-based agents rather than a single model. These agents may specialize by role, persona, evaluation dimension, or verification function; exchange messages under a protocol; and aggregate their outputs into a final score, ranking, or verdict. A common formalization writes the mapping as

$$MAM(q,o)=A\bigl(J_1^\pi(q,o),\dots,J_k^\pi(q,o)\bigr),$$

where $J_i^\pi$ is agent $i$ after interaction under protocol $\pi$, and $A$ is the aggregator. A more agentic formulation represents the judge as $\mathcal{J}=(A,T,M,C,\Phi)$, where the components are agents, a toolset, memory modules, a communication protocol, and an aggregation function (Yu, 5 Aug 2025, You et al., 8 Jan 2026). Across recent work, MAM has been used for interleaved image–text tool planning, machine translation evaluation, personalized judge construction for natural language generation, meta-judging of LLM judgments, safety auditing, and multi-turn consumer-assistant evaluation (Liu et al., 31 Mar 2026, Feng et al., 2024, Cao et al., 1 Apr 2025, Li et al., 23 Apr 2025, Lin et al., 9 Nov 2025, Herrera et al., 3 Mar 2026).

1. Formal framework and conceptual boundaries

The core distinction between MAM and single-agent judging is that evaluation is decomposed into multiple interacting judgments. In the broad survey literature, this decomposition is organized along the dimensions of Robustness, Verification, and Granularity, and along three developmental stages: Procedural, Reactive, and Self-Evolving (You et al., 8 Jan 2026). Procedural systems use predefined workflows or debate protocols; Reactive systems route among tools or sub-agents conditional on intermediate feedback; Self-Evolving systems refine rubrics and memory during evaluation.

Two topologies recur across the literature. The first is Collective Consensus, in which horizontally organized agents debate, critique, or vote until an arbiter or aggregator produces a consensus. The second is Task Decomposition, in which specialized agents handle rubric creation, evidence gathering, scoring, synthesis, or verification before a higher-level module fuses their outputs (You et al., 8 Jan 2026). The survey formulation also makes explicit that each agent can implement a policy

$$\pi_i:(\text{Task},M_i)\longrightarrow(\text{Action},\Delta M_i),$$

so a judgment step may consist of a subtask plan, a tool call, a memory read or write, or a message sent under the communication protocol $C$ (You et al., 8 Jan 2026).

This formal framing helps distinguish MAM from simple ensembling. Majority voting over independent judges is only one special case. Much of the recent literature instead studies structured communication: debate, adversarial critique, collaborative refinement, meta-judging, and localized auditing of reasoning branches (Yu, 5 Aug 2025, Yang et al., 10 Feb 2026). This suggests that MAM is best understood as a family of coordination mechanisms for evaluation, rather than as a single architecture.

2. Architectural patterns

Recent systems instantiate MAM through several recurring architectural motifs. In ATP-Bench, the term Multi-Agent MLLM-as-a-Judge (MAM) refers to a three-agent sequential evaluator for interleaved generation. The Precision Inspector checks whether each tool call is necessary and correctly specified; the Recall Inspector identifies missed opportunities for tool use; and the Chief Judge synthesizes the two reports into a final score from 0–100 (Liu et al., 31 Mar 2026). This architecture explicitly separates false positives from false negatives in tool planning.

In MAJ-EVAL, the architecture begins before scoring. Relevant domain documents are used to extract stakeholder tuples and construct detailed personas, after which persona-conditioned agents engage in in-group debate and aggregate multi-dimensional scores and feedback (Chen et al., 28 Jul 2025). Here the MAM pipeline is not only multi-agent but also multi-perspective by construction.

In CollabEval, the process is organized into three phases: initial evaluation, multi-round discussion, and final judgment. Agents first evaluate independently, then iteratively refine their assessments by attending to one another’s scores and justifications, and finally defer unresolved cases to a designated final judge (Qian et al., 1 Mar 2026). This design differs from adversarial debate by emphasizing collaborative refinement and strategic consensus checking.

In M-MAD for machine translation, the architecture is dimension-partitioned. Evaluation is decoupled into Accuracy, Fluency, Style, and Terminology; each dimension has an initial evaluator, a two-agent pro-con debate, and a final synthesis judge that merges agreed annotations into an MQM-style output (Feng et al., 2024). In the personalized Multi-Agent LLM Judge framework, the architecture instead forms an optimization loop around prompt design: a Sample Selection Agent picks diverse few-shot examples, an Evaluation Agent scores the current prompt and produces feedback, and a ReWrite Agent revises the prompt until a target threshold is met or the iteration budget is exhausted (Cao et al., 1 Apr 2025).

The resulting design space is heterogeneous but structured.

Framework	Core structure	Primary function
ATP-Bench MAM (Liu et al., 31 Mar 2026)	Precision Inspector → Recall Inspector → Chief Judge	Reference-free tool-planning evaluation
MAJ-EVAL (Chen et al., 28 Jul 2025)	Persona construction → in-group debate → aggregation	Multi-dimensional stakeholder-aligned evaluation
CollabEval (Qian et al., 1 Mar 2026)	Initial evaluation → discussion → final judge	Consensus-oriented collaborative judging
M-MAD (Feng et al., 2024)	Dimension partition → pro-con debate → synthesis	Fine-grained MT evaluation
Multi-Agent LLM Judge (Cao et al., 1 Apr 2025)	Sample selection → evaluation → rewrite loop	Automatic personalized judge design

These architectures show that MAM systems can be sequential, iterative, hierarchical, or debate-based, and may optimize a judge prompt, score a trajectory, or adjudicate multimodal tool-use plans.

3. Protocols, scoring, and aggregation

Because MAM decomposes judging into interacting components, protocol design is central. Debate-based systems typically alternate critique and response. A standard three-agent pattern uses a judge, a critic, and a defender, with the judge revising its score after reading both arguments (Yu, 5 Aug 2025). MAJ-EVAL uses a moderator to select speakers during free debate and an aggregator to merge the resulting history into final outputs (Chen et al., 28 Jul 2025). CollabEval updates agents in randomly shuffled order during each discussion round and halts early when a consensus threshold is satisfied (Qian et al., 1 Mar 2026).

Aggregation likewise varies by task. Weighted averaging, majority voting, and panel discussion all appear in the meta-judge literature. In the three-stage meta-judge selection pipeline, candidate judgments are scored on seven rubric dimensions—Accuracy of Judgment, Logical Soundness, Completeness, Fairness, Relevance to Context, Clarity of Explanation, and Impactfulness—with weights $0.20$, $0.20$, $J_i^\pi$0, $J_i^\pi$1, $J_i^\pi$2, $J_i^\pi$3, and $J_i^\pi$4, respectively; a threshold $J_i^\pi$5 is then used to filter out low-scoring judgments (Li et al., 23 Apr 2025).

Task-specific systems often introduce specialized metrics. ATP-Bench formalizes

$J_i^\pi$6

where TP is the number of necessary and correctly specified tool calls, FP is the number of superfluous or malformed calls, and FN is the number of missed opportunities. Its Chief Judge then maps the precision and recall reports into a five-tier final score rubric (Liu et al., 31 Mar 2026). M-MAD uses an MQM-style score

$J_i^\pi$7

so the final judgment is explicitly severity-sensitive (Feng et al., 2024).

Trajectory-oriented MAM systems score multi-turn interaction traces rather than isolated outputs. In the consumer-assistant setting, the full trajectory is decomposed into four weighted domains,

$J_i^\pi$8

with weights $J_i^\pi$9, $i$0, $i$1, and $i$2. Each domain is defined by binary checks, and the overall trajectory score is a weighted average of domain scores (Herrera et al., 3 Mar 2026). This formulation makes explicit that MAM can operate over traces, tool calls, and latent coordination behavior, not only over final text.

4. Empirical results and application domains

Empirical work now spans multimodal planning, NLG evaluation, MT evaluation, safety auditing, and production assistants. In ATP-Bench, ten state-of-the-art MLLMs were evaluated on 7,702 queries, including 1,592 VQA pairs, across eight categories and 25 visual-critical intents. Under MAM, Gemini 3 Pro obtained avg. FS = 79.88, SR = 81.77%, and MI = 0.49 per query, while tier-2 systems such as Claude 4.5, Grok 4.1, and GPT-5 reached FS of approximately 67–69 but differed markedly in tool-use strategy (Liu et al., 31 Mar 2026). The same study reports that three-shot prompting produced large gains for GPT-4o, with FS +12.84, SR +18.96%, and MI –0.56, and that rankings produced by different judge backbones were highly correlated, with Spearman $i$3 (Liu et al., 31 Mar 2026).

In automatic judge design for natural language generation, the Multi-Agent LLM Judge framework improved QA correctness on Instruct-QA to AUC 0.91, compared with 0.83 for a few-shot prompt, 0.81 for Continuous-Eval, 0.79 for RAGAS, and 0.78 for the initial prompt. On STS-Benchmark it achieved Pearson correlation 0.81, compared with 0.77 for a few-shot prompt, 0.67 for the initial prompt, 0.51 for Continuous-Eval, and 0.45 for RAGAS (Cao et al., 1 Apr 2025). The same paper attributes the gains to a longer, more detailed final prompt with explicit conditional instructions and curated examples.

In machine translation, M-MAD was evaluated on WMT 2023 Metrics Shared Task MQM annotations for ZH–EN, EN–DE, and HE–EN, covering 45 MT systems and 68,130 segments. On the ZH–EN test set, M-MAD achieved meta = 0.808, system Pearson = 0.987, and segment-level Acc-t = 0.517, outperforming GEMBA-MQM and EA-Prompt and rivaling reference-based automatic metrics such as XCOMET-QE-Ensemble and MetricX-23 (Feng et al., 2024). Its ablations show that removing Stage 1, the dimension partition, causes the largest degradation.

In multi-turn shopping assistants, calibrated LLM judging raised human–Judge agreement on held-out traces from 84.1% to 91.4%, and overall agreement from 88.5% to 93.5%. On 238 held-out real trajectories, Sub-agent GEPA achieved a 77.1% total pass rate, whereas MAMuT GEPA achieved 84.7%, with the largest gains in Safety (+12%) and Conversational Quality (+8%) (Herrera et al., 3 Mar 2026). This result is specifically about MAM used as an optimization target over full trajectories.

In meta-judging of LLM judgments, the three-stage pipeline based on rubric design, three advanced LLM agents, and threshold filtering improved precision by about 15.55% compared to raw judgments and about 8.37% over the single-agent baseline on JudgeBench (Li et al., 23 Apr 2025). The detailed results report raw GPT-4o-mini judgment precision of about 61.7%, a best single-agent meta-judge at about 69.3%, and a multi-agent majority-voting configuration at about 77.3% (Li et al., 23 Apr 2025).

In safety evaluation, a small-language-model multi-agent debate framework reached $i$4 on HAJailBench with a cost/query of $i$5 dollars and CostRatio 0.46 relative to GPT-4o. The same study finds that three rounds of debate yield the best balance between agreement and efficiency, improving over one round and five rounds (Lin et al., 9 Nov 2025).

5. Bias, reliability, and failure analysis

A central misconception is that adding agents necessarily reduces prejudice. The most direct evidence against that view comes from "Judging with Many Minds: Do More Perspectives Mean Less Prejudice?" (2505.19477). That study analyzes position bias, verbosity bias, chain-of-thought bias, and bandwagon bias across Multi-Agent-Debate and LLM-as-Meta-Judge. Its main result is that the debate framework amplifies biases sharply after the initial debate and sustains that increased bias in subsequent rounds, whereas meta-judge approaches exhibit greater resistance; adding PINE as a bias-free agent reduces biases in debate settings but provides less benefit in meta-judge scenarios (2505.19477).

Other work diagnoses different reliability failures. The meta-judge study reports that two-agent panels can outperform three-agent panels owing to reduced groupthink, and that aggregation strategy is task-dependent: weighted average works best on some reasoning tasks, majority voting on general tasks, and panel discussion on math (Li et al., 23 Apr 2025). "The Confident Liar" shows a substantial role asymmetry in two-agent debate: confidence aligns with judged reasoning quality roughly twice as strongly for the Constructor as for the Auditor, and critical reasoning failures are more reliably detected from confidence for the Constructor, with AUROC 0.804 versus 0.634 for the Auditor (Keramati et al., 9 Jun 2026). These findings indicate that agent role is not merely stylistic; it changes the relationship between internal confidence, judged reasoning quality, and endpoint correctness.

A separate failure mode is confabulation consensus, where agents share correlated biases and converge on the same incorrect rationale. AgentAuditor addresses this by building a Reasoning Tree, auditing localized divergence points instead of applying majority vote, and training the adjudicator with Anti-Consensus Preference Optimization (ACPO) on majority-failure cases. Across five multi-agent settings, AgentAuditor yields up to 5% absolute accuracy improvement over majority vote and up to 3% over using LLM-as-Judge (Yang et al., 10 Feb 2026). This suggests that some apparent agreement in MAM systems is evidentially weak and should itself be audited.

The broader surveys add system-level constraints: multi-round coordination and tool use increase computational cost and latency; memory introduces privacy risk; and tool-enabled agents enlarge the attack surface for prompt injection or unintended side effects (You et al., 8 Jan 2026, Yu, 5 Aug 2025). Taken together, these findings frame reliability in MAM as a problem of protocol design, bias management, and verification architecture rather than merely of model scale.

6. Open problems and research trajectory

Current research points toward more agentic, verifiable, and adaptive MAM systems. The survey literature emphasizes tool-using critics, shared memory, human–agent hybrid oversight, context-aware rubric generation, and joint multi-agent coordination objectives as core next steps (You et al., 8 Jan 2026, Yu, 5 Aug 2025). ATP-Bench proposes refining the Chief Judge’s rubric into a continuous differentiable scoring function, expanding the toolkit beyond Reference, Diffusion, Search, Edit, and Code, and incorporating human-in-the-loop feedback to calibrate agent boundaries as tool capabilities evolve (Liu et al., 31 Mar 2026).

Efficiency is another active direction. Adaptive stability detection models judge-consensus dynamics with a time-varying Beta-Binomial mixture and stops debate when successive round distributions become sufficiently similar under a Kolmogorov–Smirnov criterion. On six datasets, this yields 1–5 point accuracy gains over single judging and simple majority vote, while halting after 4–6 rounds on average and saving about 40% of LLM calls relative to fixed-length ten-round debate (Hu et al., 14 Oct 2025). In a different efficiency regime, the safety-evaluation work shows that structured, value-aligned debate allows SLMs to approach GPT-4o-level agreement at substantially lower cost (Lin et al., 9 Nov 2025).

The literature also points toward richer evaluator construction. MAJ-EVAL replaces manually crafted personas with personas automatically extracted from domain texts, while the personalized Multi-Agent LLM Judge replaces fixed evaluation prompts with an iterative loop over example selection, evaluation, and rewriting (Chen et al., 28 Jul 2025, Cao et al., 1 Apr 2025). A plausible implication is that future MAM systems will increasingly synthesize their own rubrics, personas, and stopping rules from task context rather than rely on hand-written evaluator specifications.

As a research object, MAM now sits at the intersection of evaluation methodology, agent coordination, multimodal planning, and AI auditing. The accumulated evidence does not support a single universal design. Debate can improve accuracy but amplify bias; collaboration can improve stability but risk groupthink; majority voting is simple but discards evidential structure; and more sophisticated auditors can outperform both voting and naïve judging. The field is therefore moving toward MAM systems that are not merely multi-agent, but explicitly calibrated, role-aware, structure-aware, and verification-aware (2505.19477, Yang et al., 10 Feb 2026, You et al., 8 Jan 2026).