Skip to content

Debating the Unspoken: Role-Anchored Multi-Agent Reasoning for Half-Truth Detection

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/tangyixuan/RADAR
Area: LLM Reasoning
Keywords: Half-Truth Detection, Multi-Agent Debate, Context Omission, Role-Anchored Reasoning, Adaptive Early Stopping

TL;DR

To tackle subtle half-truths where statements are factually correct in isolation yet deeply misleading due to omitted context, RADAR establishes a role-anchored multi-agent debate between stance-neutral Politician and Scientist agents over shared noisy retrieved evidence, orchestrated by a dual-threshold early-stopping judge to uncover unstated assumptions with minimal inference cost.

Background & Motivation

The rapid proliferation of online misinformation demands automated fact verification systems capable of discerning increasingly sophisticated forms of rhetorical manipulation. While modern large language models demonstrate formidable capabilities in detecting explicit falsehoods that directly contradict external evidence, they frequently stumble against a much more pervasive and elusive vulnerability: the half-truth. Half-truths are claims that remain technically accurate on the surface yet produce a severely distorted interpretation through deliberate context omission. For instance, declaring that an administration reduced the national debt by 15% is factually sound in isolation, but profoundly deceptive if it systematically omits that this decline merely followed an unmentioned 20% spike during the exact same term.

Detecting omission-based manipulation departs fundamentally from conventional fact verification because it requires adjudicating whether the available context is insufficient to justify a claim's implied conclusion, reasoning over latent presuppositions rather than surface-level contradictions. This challenge is acutely exacerbated under realistic retrieval conditions where systems operate over noisy, incomplete, and fragmented web evidence. When dense retrievers return snippets that exhibit strong lexical or topical overlap with a claim, standard single-agent verifiers and chain-of-thought (CoT) prompting readily fall prey to confirmation bias and error propagation, mistakenly validating omission-distorted claims as fully truthful. Meanwhile, existing multi-agent debate (MAD) frameworks typically assign rigid, binary pro/con stances and lean heavily on ungrounded parametric model memory, rendering them fundamentally ill-suited for examining missing background context where truth and deception coexist.

To bridge this critical gap, this work reformulates omission-aware verification as a deliberative, retrieval-grounded deliberation process driven by complementary domain expertise rather than artificial polarities. Core idea: Formulate half-truth detection as a role-anchored multi-agent debate (RADAR) between stance-neutral Politician and Scientist agents grounded on identical noisy retrieved evidence, mediated by a dual-threshold early termination controller that adaptively halts debate upon reaching sufficient evidential confidence to expose unstated context at low computational overhead.

Method

Overall Architecture

RADAR shifts fact verification from static, single-pass classification into dynamic, multi-agent adversarial deliberation under noisy information constraints. Given an input claim \(c\), the pipeline first retrieves a shared pool of top-\(m\) passages \(\mathcal{E}\) from an open corpus without assuming gold completeness. Two debating agents—a Politician focusing on public narrative and policy coherence, and a Scientist probing empirical rigor and contextual boundaries—engage in a structured multi-round dialogue constrained strictly to \(\mathcal{E}\). A neutral Judge monitors each turn using a dual-threshold early termination controller to dynamically stop the exchange when confidence suffices, culminating in a structured verdict with cited reasoning paths.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Claim c"] --> B["Context Construction under Noisy Retrieval<br/>Dense retriever compiles shared evidence pool E"]
    B --> C["Role-Grounded Multi-Round Debate<br/>Politician narrative lens vs Scientist empirical rigor"]
    C --> D["Dual-Threshold Early Termination Controller<br/>Judge evaluates stop margin s and veracity confidence c"]
    D -->|s < tau_s or c < tau_v and round < max| C
    D -->|Early stop triggered or max rounds reached| E["Structured Verdict Output<br/>[REASON] evidence trace + [VERDICT] label"]

Key Designs

1. Context Construction under Noisy Retrieval: Eliminating Gold Evidence Assumptions via Shared Grounding Conventional omission benchmarks (such as TRACER) presume idealized access to human-curated gold evidence containing pre-annotated hidden context, a setting unachievable in real-world deployments. RADAR dismantles this unrealistic reliance by introducing an end-to-end noisy retrieval pipeline. For any input claim \(c\), a dense retriever (bge-base-en-v1.5) retrieves the top-\(m\) passages from an uncurated background corpus of 41,952 documents to form the shared pool \(\mathcal{E} = \text{RETRIEVE}(c)\). This evidence inevitably includes irrelevant passages, fragmented observations, and partial truths. Crucially, \(\mathcal{E}\) serves as the strict, identical grounding anchor for all debating agents. Forcing agents to cite specific snippets in \(\mathcal{E}\) curbs ungrounded parametric hallucinations and eliminates information asymmetry, ensuring that dialectical disagreements stem purely from reasoning and critical interpretation rather than divergent data access.

2. Role-Grounded Multi-Round Debate: Complementary Expertise Personas Replacing Binary Stances Unlike conventional multi-agent debate architectures that force agents into artificial "Advocate" versus "Critic" polarities, RADAR assigns agents complementary, domain-aligned cognitive roles without predefined pro/con positions. The Politician agent evaluates claims through public-interest rhetoric, stakeholder framing, and policy-communication coherence, naturally uncovering the societal narrative and communicative intent behind a statement. Conversely, the Scientist agent acts as an empirical methodologist, scrutinizing data quality, baseline validity, selective framing, and temporal or causal inconsistencies. Across Opening, Rebuttal, and Closing rounds, neither agent is coerced to blindly defend or attack the claim; instead, the Politician surfaces the plausible framing while the Scientist highlights what critical conditions or baseline numbers have been silently bypassed. This natural tension allows the unspoken omissions of half-truths to be methodically surfaced through adversarial dialogue.

3. Dual-Threshold Early Termination Controller: Dynamic Efficiency and Reasoning Regularization While iterative debate uncovers hidden nuances, exhaustive multi-round exchanges impose heavy latency and token costs, and risk introducing semantic drift or excessive nitpicking on straightforward claims. RADAR integrates a training-free, dual-threshold early stopping controller evaluated by the Judge at the conclusion of each round. The controller first monitors the stop margin: $\(s = p(\text{STOP}) - p(\text{CONTINUE})\)$ derived from the normalized softmax probability over decision tokens. Debate continues if \(s < \tau_s\). To guard against premature termination on subtle half-truths where initial agreement masks deep omissions, the controller additionally requires the Judge's veracity prediction confidence to meet a threshold: $\(c = \max_{y \in \mathcal{Y}} p(y)\)$ where \(\mathcal{Y} = \{\text{True}, \text{Half-True}, \text{False}\}\). Early termination is triggered if and only if both conditions hold (\(s \ge \tau_s\) and \(c \ge \tau_v\)). Thresholds are calibrated offline on a validation set using cached logits from a single forward pass, providing dynamic compute allocation that prunes 15% to 37% of token overhead while boosting veracity accuracy.

Benchmark & Evaluation Protocol

RADAR requires no parameter fine-tuning. Experiments are conducted across two major benchmarks: - POLITIFACT-HIDDEN: Comprises ~15,000 real-world political claims annotated for omission-based verification across three classes (True, Half-True, False). Split into 11,994 train, 1,000 dev, and 2,000 test claims, backed by a corpus of 41,952 unstructured articles. - AVeriTeC: A challenging open-web real-world verification benchmark featuring multi-step question-answering evidence. "Conflicting Evidence/Cherry-picking" is mapped to Half-True, retaining Not Enough Evidence (NEI) alongside True and False in a four-class generalization setup. - Backbone & Retrieval Settings: Tested under both Full Gold Evidence and Realistic Retrieved Evidence (dense retrieval top-20). Evaluated across diverse LLM backbones: GPT-4o-mini, LLaMA3-8B-Instruct, and Qwen2.5-7B-Instruct.

Key Experimental Results

Main Results

Performance comparison under the realistic Retrieved Evidence setting (Top-20 retrieved passages) on the POLITIFACT-HIDDEN test set:

Backbone Model Accuracy (%) Macro-F1 (%) F1_True (%) F1_Half-True (%) F1_False (%)
GPT-4o-mini FIRE 60.3 46.9 30.7 34.1 75.9
GPT-4o-mini D2D (Binary Debate) 63.0 50.9 36.3 39.7 76.5
GPT-4o-mini RADAR_single (Single Agent) 58.4 51.0 41.2 41.5 70.2
GPT-4o-mini RADAR_multi (Ours) 77.7 (+19.3) 63.3 (+12.3) 45.6 (+4.4) 56.5 (+15.0) 87.6 (+17.4)
LLaMA3-8B FIRE 50.4 43.4 30.9 39.7 59.6
LLaMA3-8B D2D (Binary Debate) 64.1 45.1 21.7 34.1 79.5
LLaMA3-8B RADAR_single (Single Agent) 57.3 52.0 44.4 44.8 66.6
LLaMA3-8B RADAR_multi (Ours) 77.3 (+20.0) 63.2 (+11.2) 46.3 (+1.9) 56.7 (+11.9) 86.5 (+19.9)
Qwen2.5-7B FIRE 63.9 46.9 29.2 32.1 79.5
Qwen2.5-7B D2D (Binary Debate) 69.0 48.5 22.2 41.2 82.1
Qwen2.5-7B RADAR_single (Single Agent) 71.2 54.9 39.4 41.1 84.3
Qwen2.5-7B RADAR_multi (Ours) 76.7 (+5.5) 63.1 (+8.2) 48.2 (+8.8) 54.9 (+13.8) 86.2 (+1.9)

Note: Parenthetical gains reflect the improvement of RADAR_multi over its single-agent counterpart RADAR_single under identical retrieval conditions.

Cross-domain four-class generalization on AVeriTeC (Retrieved Evidence setting):

Model Accuracy (%) Macro-F1 (%) F1_True (%) F1_Half-True (%) F1_False (%) F1_NEI (%)
FIRE 26.8 21.1 11.9 14.7 45.6 12.1
D2D 63.8 46.9 55.3 16.9 80.4 35.1
RADAR_multi 77.8 57.7 79.6 20.3 86.8 43.9

Ablation Study & In-depth Analysis

Impact of Adaptive Early Stopping on verification performance and generation token efficiency:

Backbone Configuration Accuracy (%) Macro-F1 (%) Avg Tokens / Claim Relative Compute
LLaMA3-8B-Instruct D2D 64.1 45.1 2483.3 100.0%
LLaMA3-8B-Instruct FIRE 50.4 43.4 754.5 30.4%
LLaMA3-8B-Instruct RADAR (Fixed 3 Rounds) 77.2 62.3 1834.7 73.9%
LLaMA3-8B-Instruct RADAR (Early Stopping) 77.3 63.2 1551.4 62.5% (-15.4%)
Qwen2.5-7B-Instruct D2D 69.0 48.5 3686.9 100.0%
Qwen2.5-7B-Instruct FIRE 63.9 46.9 1026.2 27.8%
Qwen2.5-7B-Instruct RADAR (Fixed 3 Rounds) 76.5 62.6 1751.5 47.5%
Qwen2.5-7B-Instruct RADAR (Early Stopping) 76.7 63.1 1723.7 46.8% (-1.6%)

Scaling debater count with expertise-based personas (fixed LLaMA3-8B, 150 sampled claims, Judge excluded from count):

Number of Debaters Debater Role Configuration Accuracy (%) Macro-F1 (%) F1_True (%) F1_Half-True (%) F1_False (%)
1 Single Debater Self-Reasoning 61.3 61.5 53.9 60.0 70.7
2 Politician + Scientist (Default) 64.0 62.8 46.8 60.2 81.5
3 Politician + Scientist + Journalist 58.0 54.6 30.0 60.1 73.6
4 Politician + Scientist + Journalist + Domain Expert 59.3 56.4 34.9 56.7 77.7

Key Findings

  • Gains Concentrate Heavily on Omission-Sensitive Classes: Under noisy retrieval, multi-agent debate delivers massive leaps in Half-True F1 (+11.9% to +15.0%) and False F1 (+17.4% to +19.9%) compared to single-agent baselines. This confirms that deliberative cross-examination is indispensable for detecting when superficially accurate statements are crippled by missing background.
  • Expertise Personas Outperform Position-Based Polarization: Replacing generic Advocate vs. Critic debaters with the Politician vs. Scientist pair elevates Macro-F1 from 56.4% to 62.3% under noisy retrieval. Position-constrained agents often engage in ungrounded rhetorical obstinacy, whereas complementary expertise drives natural alignment between communicative narrative and empirical scrutiny.
  • Two Debaters Strike the Optimal Trade-off: Scaling agent counts beyond two yields diminishing returns. Expanding to 3 or 4 debaters degrades accuracy to 58.0% and collapses F1_True to 30.0% due to redundant noise, argument dilution, and conflicting claims. A compact two-agent core provides the cleanest dialectic tension.
  • Early Stopping Simultaneously Cuts Cost and Boosts Accuracy: In LLaMA3-8B, the adaptive controller prunes 15.4% of total tokens while increasing Macro-F1 from 62.3% to 63.2%. By terminating straightforward claims early (846 out of 2,000 cases stopped at Round 1), it shields easy instances from semantic drift and hyper-critical divergence.

Highlights & Insights

  • Reframing Verification from Disproving to Uncovering: Traditional verification assumes lies come with factual contradictions. RADAR establishes that the hardest misinformation thrives on true premises, demonstrating that multi-agent dialectic is a powerful tool to inspect "what is left unsaid."
  • Strict Evidence Grounding Prevents Hallucinatory Debates: Rather than allowing LLMs to debate from untethered internal memory, forcing agents to substantiate claims using shared noisy retrieval produces audit-traceable, fact-checked verdicts that human evaluators can inspect.
  • Lightweight Offline Threshold Calibration: The dual-threshold stopping mechanism requires zero model training and determines cutoffs using cached logits from a single pass over dev data, providing an off-the-shelf mechanism to rein in multi-agent inference inflation.

Limitations & Future Work

  • Risk of Over-Scrutiny on Truthful Statements: Adversarial critique can occasionally become overzealous. On strictly factual claims (True), hyper-critical deconstruction by the Scientist can coax the Judge into downgrading the statement to Half-True, shifting the precision-recall trade-off.
  • Static Retrieval vs. Dynamic Active Search: The evidence pool is fixed prior to debate. If debating agents identify a specific missing fact, the system cannot currently execute an active, mid-debate search query to retrieve supplementary context on demand.
  • Fragility to Fine-Grained Constraint Violations: Error analysis of 50 failure cases revealed that 56% of misclassifications stem from failing to enforce explicit numeric, temporal, or geographic boundaries in the claim, highlighting the need to integrate symbolic constraint trackers into agent reasoning loops.
  • vs TRACER (Tang et al., 2025): TRACER pioneered omission modeling through critical hidden evidence but relied on a static single-agent pipeline and required gold evidence. RADAR operates under open, noisy retrieval and discovers omitted context via dynamic multi-agent interaction.
  • vs D2D (Han et al., 2025) & TruEDebate (Liu et al., 2025): Prior MAD frameworks applied fixed pro/con roles to fake news, relying heavily on ungrounded parametric knowledge. RADAR enforces shared retrieval grounding and adopts stance-neutral expertise personas, surpassing D2D by 12.4% to 18.1% Macro-F1 on half-truth benchmarks.
  • vs Iterative Single-Agent Systems (FIRE, Xie et al., 2025): FIRE relies on iterative search-and-verify loops driven by explicit contradictions. It struggles to detect omissions where retrieved passages appear superficially supportive, lagging significantly behind RADAR.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Replaces artificial pro/con polarities with stance-neutral expertise roles grounded strictly on shared noisy retrieval to tackle context omission.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across 2 benchmarks, 3 LLM backbones, gold vs. retrieved setups, detailed stopping sensitivity, debater scaling, and manual error breakdowns.
  • Writing Quality: ⭐⭐⭐⭐⭐ Clear, structured, and coherent narrative that thoroughly unpacks the subtleties of half-truth verification.
  • Value: ⭐⭐⭐⭐⭐ Delivers an actionable, training-free, and cost-efficient paradigm for robust real-world fact-checking and multi-agent reasoning.