Two Minds are Safer Than One: Argumentative Llm Agents for Clinical Diagnosis

MedAgent-R1: Faithfulness-Aware Reinforcement Learning for Evidence-Grounded Medical Reasoning

When medical AI systems hallucinate clinical reasoning, the consequences extend beyond incorrect answers: fabricated justifications that superficially reference retrieved evidence can mislead clinicians into unsafe treatment decisions. Medical reasoning agents must therefore produce not only correct answers but also faithful justifications that clinicians can verify against cited evidence. We identify a systematic failure mode in RL-trained retrieval agents: outcome-only rewards improve accuracy while degrading faithfulness, a phenomenon we term confident hallucination. The agent learns to answer from parametric memory and backfill plausible but unsupported justifications; citation fabrication rates rise from 16.5% to 31.8% even as accuracy improves by 5 points over the supervised baseline. We address this with a faithfulness-gated reward design: accuracy credit is conditioned on evidence grounding via a hard gate, complemented by retrieval validity and conciseness signals that close exploitation paths unique to agentic retrieval. The resulting system, MedAgent-R1, reduces citation fabrication from 31.8% to 4.7% and raises evidence completeness from 58.7 to 82.6 while maintaining 75.1% accuracy, with 13.2-point gains on HealthBench Safety. Under the same agentic retrieval setup, MedAgent-R1 outscores GPT-4o on faithfulness-specific dimensions (Factual Support 4.55 vs. 4.25; Overclaiming 4.40 vs. 4.15) while remaining below GPT-4o in overall accuracy, suggesting that explicit faithfulness training yields evidence-grounding gains not achieved by scaling alone.

Right for the Wrong Reasons: A Benchmark for Hallucination and Clinical Safety in AI Health Triage

EVADE: Evidence-Verified Agentic Diagnosis with Escape

Medical vision-language models (VLMs) can achieve high accuracy but remain unreliable: they are systematically overconfident, benefit little from test-time reasoning, and lack the ability to reliably calibrate trust in their own responses. We introduce EVADE (Evidence-Verified Agentic Diagnosis with Escape), an inferential, non-training method that enhances the safety of deploying a single frozen VLM. EVADE responds and, when uncertain, localises the region most diagnostically relevant, re-answers on a zoomed view, and commits only when both the entire image and the zoomed view responses agree; otherwise, it abstains. To directly address verification hallucination in single-model self-checking, our main idea is to verify gate consistency across different image views rather than re-reading the model's own text. Experimental evaluation on VQA-RAD, SLAKE, and PathVQA using Qwen2.5-VL-7B reports that EVADE is the only method that simultaneously improves both calibration and selective risk while maintaining accuracy, reducing expected calibration error (ECE) by up to 45% compared to zero-shot. Chain-of-thought, self-consistency, and self-verification all fail at least one axis. A grounding analysis reports that self-proposed regions perform better at diagnostic structure localisation than centres or random crops. However, a 7B VLM cannot use this localisation to revise answers. Therefore, reliability gains come from the consistency gate and calibrated abstention.

CSMAD: Hallucination Detection via Multi-Agent Debate with NLI-Verified Contradictory Statements

Large Language Models (LLMs) are prone to hallucinations, producing fluent but factually incorrect statements. Recent multi-agent debate methods improve hallucination detection by jointly improving reasoning and decision-making. However, existing approaches either collaborate which amplifies shared overconfidence, or adopt adversarial preset stances, that can inject incorrect information complicating decision making. To address this, we propose Contradictory Statement Multi-Agent Debate (CSMAD), a multi-agent framework that creates structured disagreement by generating a contradictory claim for each input claim. CSMAD asks independent agents to evaluate the claim and the contradictory claim, which encourages different lines of reasoning without assigning preset stances. When the outcome is non-discriminative; both the contradictory statements are either accepted or rejected; the agents exchange rationales and update their judgments after considering opposing evidence. A final judge then decides the truth of the original claim, using both arguments as context. To make contradictory statement generation reliable, we add a Natural Language Inference (NLI) based verifier that checks whether the generated statement actually contradicts the original claim; if it does not, the system falls back to an explicit negation-based contradiction. Across public benchmarks for question answering and scientific claim verification, as well as a proprietary e-commerce claims dataset, we show that CSMAD consistently outperforms the strongest baseline for both large (Claude-3.5 Sonnet) and medium-sized (Qwen3-8B) language models, improving F1 by +2.3 and +4.1 points, respectively, while reducing LLM token cost by 28%.

Self-Reflective Large Language Models for Reducing AI Hallucinations: A Novel Framework for Reliable Generative AI

- Large Language Models (LLMs) have become a central technology for generative artificial intelligence, but their tendency to produce fluent yet factually unsupported information remains a major barrier to dependable deployment. This paper proposes a self-reflective framework in which an LLM generates an answer, identifies claims that may be uncertain, performs an internal verification stage, and revises the response before delivery. Unlike a single-pass generation process, the proposed approach separates generation, claim inspection, evidence-oriented verification, and response refinement. The framework is designed to reduce unsupported claims while preserving useful information and acceptable response latency. The paper presents a research-oriented evaluation methodology using factuality, unsupported-claim rate, answer completeness, calibration, and computational overhead as evaluation dimensions. The proposed framework can be integrated with retrieval-augmented generation, external knowledge sources, or domain-specific validation modules. The study argues that self-reflection should be treated not merely as prompt engineering but as a structured reliability layer for generative AI systems.

Evidence-Bounded Mental Health Reasoning from Heterogeneous Speech Protocols

The Evidence Package Benchmark is introduced, integrating 1,870 packages across six heterogeneous sources with explicit modality masks and evidence permissions, and EviBound, a protocol-aware evidence control framework is proposed, a protocol-aware evidence control framework for safer clinical NLP research.