MARCH: Multi-Agent Reinforced Self-Check

• The paper introduces MARCH, a framework that reduces LLM hallucinations by decomposing answers into atomic claim-level units and verifying them with deliberate information asymmetry.
• It employs a three-agent architecture where a Solver generates responses, a Proposer extracts verifiable QA pairs, and a Checker verifies claims solely from retrieved documents.
• Empirical results show that an 8B LLM using MARCH significantly outperforms standard RAG approaches on factuality and multi-hop QA benchmarks.

Multi-Agent Reinforced Self-Check for Hallucination (MARCH) is a reinforcement-learned, multi-agent framework for reducing LLM hallucinations in Retrieval-Augmented Generation (RAG) by decomposing answers into claim-level units and verifying them under deliberate information asymmetry. In its canonical formulation, a Solver generates an initial response, a Proposer converts that response into verifiable atomic propositions, and a Checker re-answers those propositions from retrieved documents alone, without access to the Solver’s original text. This design targets the confirmation bias that affects conventional LLM-as-a-judge pipelines and supports joint multi-agent reinforcement learning with a strict factuality reward (Li et al., 25 Mar 2026).

1. Terminology, scope, and problem setting

MARCH addresses a specific failure mode in RAG: context-conflicting hallucinations such as wrong numbers, distorted timelines, and incorrect aggregations that appear even when the necessary information is present in retrieved documents. The central diagnosis is that post-hoc verification by an LLM that sees the query, the retrieved documents, and the model’s answer is susceptible to confirmation bias, because the verifier can become anchored to the answer it is asked to evaluate. MARCH reformulates factuality as a claim-level verification problem in which the verifier is intentionally denied the generator’s final response (Li et al., 25 Mar 2026).

The framework is defined over a user query $x$, retrieved documents $D = [d_1, d_2, \dots, d_l]$, and a policy $\pi_\theta$ that outputs a response $y$. The paper first states an ideal RLVR objective that would compare extracted factual statements $A(y)=\{a_1,\dots,a_n\}$ against unavailable oracle annotations $A^*=\{a_1^*,\dots,a_n^*\}$:

$$\max_\theta \; \mathbb{E}_{x\sim\cdot,\; y\sim \pi_\theta(\cdot\mid x,D)} \left[R(\{a_i,a_i^*\}_{i=1}^n)\right].$$

Because $A^*$ is unavailable in realistic settings, MARCH replaces oracle answers with an internally generated, document-grounded reference:

$$\max_\theta \; \mathbb{E}_{x\sim\cdot,\; y\sim \pi_\theta(\cdot\mid x,D)} \left[R(\{(q_i,\hat a_i)\}_{i=1}^n)\right].$$

This makes the training signal self-contained and verifiable at the level of atomic claims rather than coarse response-level plausibility (Li et al., 25 Mar 2026).

The acronym is not unique. In the radiology paper "MARCH: Multi-Agent Radiology Clinical Hierarchy for CT Report Generation" it denotes Multi-Agent Radiology Clinical Hierarchy, and that system explicitly states that there is no reinforcement learning in its design (Lin et al., 17 Apr 2026). In later literature, however, "MARCH" is also used more broadly as a conceptual label for multi-agent self-check pipelines beyond the original RAG setting, including medical auditing and safety-oriented verification systems (Osama et al., 12 Jun 2026).

2. Agent architecture and deliberate information asymmetry

MARCH uses three roles parameterized by the same base policy $\pi_\theta$ but separated by system prompts and information access. The architecture is intentionally asymmetric: decomposition and verification are not allowed to share the same full context as generation.

Agent	Input	Function
Solver	$D = [d_1, d_2, \dots, d_l]$0	Generates the initial RAG answer $D = [d_1, d_2, \dots, d_l]$1
Proposer	$D = [d_1, d_2, \dots, d_l]$2	Converts $D = [d_1, d_2, \dots, d_l]$3 into claim-level QA pairs $D = [d_1, d_2, \dots, d_l]$4
Checker	$D = [d_1, d_2, \dots, d_l]$5	Re-answers each $D = [d_1, d_2, \dots, d_l]$6 from documents alone

The Solver is defined as $D = [d_1, d_2, \dots, d_l]$7 and is prompted to respond strictly from the provided passages. The Proposer is defined as $D = [d_1, d_2, \dots, d_l]$8 and atomizes the response into verifiable QA pairs $D = [d_1, d_2, \dots, d_l]$9. The Checker is defined as $\pi_\theta$0 and produces grounded answers $\pi_\theta$1 from the retrieved corpus alone (Li et al., 25 Mar 2026).

A defining constraint is that the Checker never sees the Solver’s response $\pi_\theta$2 or the Proposer’s extracted answers $\pi_\theta$3. It only receives the questions $\pi_\theta$4 and the retrieved documents $\pi_\theta$5. The paper identifies this as deliberate information asymmetry, intended to break the self-confirmation loop that arises when a verifier is exposed to the generator’s wording or numeric claims. The Proposer, conversely, sees $\pi_\theta$6 but does not see $\pi_\theta$7, which isolates decomposition quality from retrieval-grounded verification (Li et al., 25 Mar 2026).

In the reported instantiation, atomic propositions are numerically grounded. Each question $\pi_\theta$8 is crafted so that the correct answer is a single number $\pi_\theta$9 present in $y$0, and prompt rules require specific questions and pure numeral answers with no percent signs, ranges, or words. The Checker answers each question with evidence and either a number or "Cannot answer." This numerical restriction provides an exact-match verification interface with low ambiguity (Li et al., 25 Mar 2026).

3. Reinforcement learning formulation and zero-tolerance reward

The full constrained objective is written as

$y$1

subject to

$y$2

$y$3

The reward is a Zero-Tolerance Reward:

$y$4

This all-or-nothing formulation encodes the claim that a single hallucinated atomic fact invalidates the entire response (Li et al., 25 Mar 2026).

Training uses Proximal Policy Optimization. The same $y$5 plays all three roles, and both the Solver trajectory $y$6 and the Checker trajectory $y$7 receive the same terminal reward. Advantages are computed separately with GAE, and updates are regularized with a KL penalty to a frozen reference model $y$8. The joint gradient is

$y$9

Here $A(y)=\{a_1,\dots,a_n\}$0 abbreviates $A(y)=\{a_1,\dots,a_n\}$1 for the Solver and $A(y)=\{a_1,\dots,a_n\}$2 abbreviates $A(y)=\{a_1,\dots,a_n\}$3 for the Checker (Li et al., 25 Mar 2026).

Several implementation details are central to the training dynamics. The Checker is multi-sampled per question and majority-voted to reduce stochasticity. Joint Solver-plus-Checker optimization outperforms Solver-only training. The paper also reports reward hacking risk: the number of proposed questions can drop during training, so a prompt constraint requiring "at least $A(y)=\{a_1,\dots,a_n\}$4 questions" is used to stabilize coverage. Among reward variants, Zero-Tolerance Reward outperforms an error-rate reward, and a negative-penalty form outperforms a positive $A(y)=\{a_1,\dots,a_n\}$5 incentive form (Li et al., 25 Mar 2026).

4. Training setup, evaluation suites, and empirical performance

The base policy is initialized from Meta-Llama3.1-8B-Instruct. Online MARL training uses BioASQ with 4,721 samples, 2WikiMultiHopQA with 4,500 samples, and MuSiQue with 4,500 samples. The paper distinguishes MARCH-STEM, trained on BioASQ, from MARCH-General, trained on 2WikiMultiHopQA and MuSiQue. Training uses VerL, FSDP, and vLLM with temperature $A(y)=\{a_1,\dots,a_n\}$6, single-epoch PPO, global batch size $A(y)=\{a_1,\dots,a_n\}$7, maximum prompt length $A(y)=\{a_1,\dots,a_n\}$8, maximum response length $A(y)=\{a_1,\dots,a_n\}$9, actor learning rate $A^*=\{a_1^*,\dots,a_n^*\}$0, critic learning rate $A^*=\{a_1^*,\dots,a_n^*\}$1, $A^*=\{a_1^*,\dots,a_n^*\}$2, and KL coefficient $A^*=\{a_1^*,\dots,a_n^*\}$3. In evaluation, eight independent generations are sampled per test query and majority vote determines the final verdict (Li et al., 25 Mar 2026).

The evaluation covers RAGTruth, FaithBench, FACTS Grounding, ContextualJudgeBench, and multi-hop QA benchmarks including HotpotQA, 2WikiMultiHopQA, and MuSiQue. The reported results position an 8B open model with MARCH near or above several larger or proprietary baselines on factuality-oriented tasks (Li et al., 25 Mar 2026).

Evaluation suite	Base Llama3.1-8B-Instruct	MARCH
RAGTruth + FaithBench average	55.20	74.93 (STEM), 75.23 (General)
FACTS Grounding	57.09	85.23 (STEM), 80.12 (General)
ContextualJudgeBench average	29.7	52.3 (STEM), 51.6 (General)

On RAGTruth and FaithBench, the base model averages 55.20, while MARCH-STEM reaches 74.93 and MARCH-General reaches 75.23. The MARCH-STEM breakdown is Summary 86.67, Data-to-Text 72.67, QA 83.45, and FaithBench 56.94; MARCH-General reports Summary 92.67, Data-to-Text 72.00, QA 83.45, and FaithBench 52.78. On FACTS Grounding, the base score is 57.09%, MARCH-STEM reaches 85.23%, and MARCH-General reaches 80.12%. On ContextualJudgeBench, the base average is 29.7, while MARCH-General reaches 51.6 and MARCH-STEM reaches 52.3 (Li et al., 25 Mar 2026).

On multi-hop QA, the paper reports strong 8B results. Under 10-shot prompting, MARCH achieves 73.6 on HotpotQA, 40.8 on MuSiQue, and 69.4 on 2WikiMultiHopQA; under chain-of-thought prompting, it reports 70.6, 36.2, and 70.6, respectively. In ablations, joint optimization consistently exceeds Solver-only training, and Zero-Tolerance Reward exceeds Error-Rate Reward. For example, on General-domain joint training, Error-Rate Reward gives 55.46% average performance, while Zero-Tolerance Reward gives 61.25% (Li et al., 25 Mar 2026).

5. Extensions, neighboring frameworks, and broader uses of the term

Subsequent work has used "MARCH" both as the name of the original RAG framework and as a broader design pattern for multi-agent hallucination mitigation. "Trust but Verify" explicitly frames a five-agent clinical auditing system as a concrete instantiation of a general Multi-Agent Reinforced Self-Check for Hallucination framework. Its architecture uses a Router, Medical Clinical Generator, Entity Extractor, Safety Auditor, and General Chat agent, and it reduces Hallucination Error Rate from 90.29%–99.03% in vanilla settings to 35.92%–40.78% under agentic auditing, with an approximate mean HER reduction of ~53% across models (Osama et al., 12 Jun 2026).

Other work adapts the same conceptual pattern without reinforcement learning. "SEVADE" separates multi-agent reasoning from a decoupled Rationale Adjudicator and reports average improvements of 6.75% in Accuracy and 6.29% in Macro-F1 for sarcasm detection (Liu et al., 9 Aug 2025). "SelfCheckAgent" uses a Symbolic Agent, Specialized Detection Agent, and Contextual Consistency Agent for zero-resource hallucination detection; on WikiBio, its Contextual Consistency Agent with Llama 3.1 and chain-of-thought reaches 93.64 on NonFactual detection, 70.26 on Factual detection, and 78.48 on Ranking (Muhammed et al., 3 Feb 2025). "Towards Detecting LLMs Hallucination via Markov Chain-based Multi-agent Debate Framework" uses Trust, Skeptic, and Leader roles in an inference-only verification scheme and reaches 77.68 claim-level accuracy on Factool (Sun et al., 2024). "Interpreting and Mitigating Hallucination in MLLMs through Multi-agent Debate" uses self-reflection and debate rather than RL, improving GPT-4o accuracy on POPE Random from 85.77% to 87.83% (Lin et al., 2024).

A further line of work treats hallucination as a system-level collective process rather than only a per-answer verification problem. "Collective Hallucination in Multi-Agent LLMs: Modeling and Defense" models hallucination propagation over interaction graphs and introduces confidence-weighted aggregation, adaptive impact regulation, external claim verification, and selective isolation. It reports hallucination reduction by up to 39.0%, factual accuracy improvement from 0.79 to 0.87, semantic consistency improvement from 0.75 to 0.84, and adversarial amplification limited to 1.08 instead of 1.45 (Jamshidi, 6 Jun 2026). This suggests that the original MARCH emphasis on verifier isolation can be complemented by topology-aware control when hallucinations are recursively exchanged across agents.

Domain-specific systems also adopt MARCH-like self-check loops. "MedMMV" describes diversified short rollouts, a Hallucination Detector supervisor, and Combined Uncertainty scoring, with accuracy improvements of up to 12.7% on medical benchmarks (Liu et al., 29 Sep 2025). "InEx" combines entropy-guided introspection with textual and visual self-reflection agents and reports gains of 3.8%–16.2% on general-purpose benchmarks and 6.5%–26.7% on hallucination benchmarks (Yang et al., 2 Dec 2025). "EH-Benchmark" in ophthalmology instantiates a retriever-planner-specialist-validator workflow and reports a full-system improvement of +21.95% F1 on A2 instance-level logical tasks relative to GPT-4o (Pan et al., 24 Jul 2025). These systems do not reproduce the exact three-agent MARL pipeline of the original MARCH paper, but they extend its core ideas of role specialization, self-check, and verifier-guided correction.

6. Limitations, open directions, and significance

The original MARCH paper is explicit about several limitations. Its current atomic-proposition scheme focuses on numerical claims, because exact-match identity checks are easier to ground and evaluate than categorical or multi-span propositions. The Checker, although blinded, is still an LLM and can fail under high document noise or very long contexts. Retrieval quality remains a bottleneck: if the documents are insufficient or heavily contaminated by irrelevant passages, the Checker may return "Cannot answer," and the Solver is then penalized even when the answer would be recoverable from a better retriever. Reward hacking through reduced question count is another observed failure mode, mitigated only by prompt constraints such as requiring at least $A^*=\{a_1^*,\dots,a_n^*\}$4 questions (Li et al., 25 Mar 2026).

These constraints delimit the scope of what MARCH currently verifies well. A plausible implication is that the framework is strongest when claims can be decomposed into low-ambiguity units with clear grounding interfaces. Later domain-specific systems point toward extensions beyond numeric RAG facts: radiology MARCH uses stance-based consensus over anatomically grounded report revisions rather than RL (Lin et al., 17 Apr 2026), while ophthalmic and medical auditing systems use planner-specialist-validator loops, retrieval grounding, and refusal policies to control risk in clinically sensitive settings (Pan et al., 24 Jul 2025, Osama et al., 12 Jun 2026).

The broader significance of MARCH lies in its shift from monolithic answer judging to structured, asymmetric verification. In the original formulation, an 8B-parameter LLM equipped with MARCH achieves performance competitive with powerful closed-source models (Li et al., 25 Mar 2026). More generally, the subsequent literature suggests that multi-agent self-check can be realized through several mechanisms—joint MARL, deterministic auditing loops, confidence-weighted interaction control, or decoupled adjudicators—but the central design principle remains stable: hallucination mitigation improves when generation, decomposition, and verification are separated, and when the verifying pathway is protected from the generator’s own narrative.