Multi-Agent Debate (MAD) has emerged as a promising inference scaling method for Large Language Model (LLM) reasoning. However, it frequently suffers from belief entrenchment, where agents reinforce shared errors rather than correcting them. Going beyond merely identifying this failure, we decompose it into two distinct root causes: (1) the model’s biased static initial belief and (2) homogenized debate dynamics that amplify the majority view regardless of correctness. To address these sequentially, we propose DReaMAD (Diverse Reasoning via MultiAgent Debate). Our framework first rectifies the static belief via strategic prior knowledge elicitation, then reshapes the debate dynamics by enforcing perspective diversity. Validated on our new MetaNIM Arena benchmark, DReaMAD significantly mitigates entrenchment, achieving a +9.5% accuracy gain over ReAct prompting and a +19.0% higher win rate than standard MAD.

Retrieval-augmented generation (RAG) is increasingly deployed on unverified corpora, where an adversary can inject poisoned documents to steer the model toward an attacker-specified answer. Existing defenses rely on auxiliary heuristics and break down once their assumptions about the attack type or poisoning level are violated. We propose SAFER-RAG, a defense that derives its reliability signal from the adversarial objective itself. Since poisoned documents are explicitly optimized to rank highly for the target query, they remain anchored at top ranks under semantically equivalent rewrites, while clean documents shift naturally. SAFER-RAG captures this rank-stability gap through spectral analysis, partitions retrieved chunks into Safe and Unsafe Sets, and adaptively selects evidence under a safe-first strategy that avoids over-filtering useful evidence on clean or lightly poisoned corpora. Across three QA benchmarks, three attack types, and four poisoning levels, SAFER-RAG consistently achieves the highest accuracy and the lowest attack success rate. The code will be released upon acceptance.

Large language models (LLMs) are increasingly used in scientific writing, yet their citation behavior remains poorly understood. We introduce CiteSpectrum, a controlled dataset for analyzing citations across five levels of AI involvement, from LLM-Only generation to Human-Authored papers. Using citation-behavior features across fidelity, usage, and topology, we show that AI-generated papers differ from human-written papers in how citations are validated, grounded, used, and structurally organized. Even when retrieval and human guidance reduce ghost citations, AI-generated papers contain more unsupported claims, rely on front-section evidence, use citations more affirmatively, and form broader but less interconnected reference networks. Moreover, citation behavior alone enables a lightweight classifier to outperform strong text-based detectors, achieving 0.880 balanced accuracy and 0.944 AUROC. These results suggest that citation behavior provides a strong, interpretable signal of AI involvement beyond conventional textual patterns. The dataset is available at: https://github.com/DISL-Lab/CiteSpectrum-EMNLP26.

Bargaining is often regarded as a logical arena rather than an art or a matter of intuition, yet Large Language Models (LLMs) still struggle to navigate it due to limited strategic depth and difficulty adapting to complex human factors. Current benchmarks rarely capture this limitation. To bridge this gap, we present a utility feedback centric framework. Our contributions are: (i) AgoraBench, a new benchmark spanning nine challenging settings (e.g., deception, monopoly) that supports diverse strategy modeling; (ii) human-aligned, economically grounded metrics derived from utility theory. This is operationalized via agent utility, negotiation power, and acquisition ratio that implicitly measure how well the negotiation aligns with human preference and (iii) a human preference grounded dataset with learning pipeline that strengthens LLMs’ bargaining ability through both prompting and finetuning. Empirical results indicate that baseline LLM strategies often diverge from human preferences, while our mechanism substantially improves negotiation performance, yielding deeper strategic behavior and stronger opponent awareness.

Developing effective text summarizers remains a challenge due to issues like hallucinations, key information omissions, and verbosity in LLM-generated summaries. This work explores using LLM-generated feedback to improve summary quality by aligning the summaries with human preferences for faithfulness, completeness, and conciseness. We introduce FeedSum, a large-scale dataset containing multi-dimensional LLM feedback on summaries of varying quality across diverse domains. Our experiments show how feedback quality, dimensionality, and granularity influence preference learning, revealing that high-quality, multi-dimensional, fine-grained feedback significantly improves summary generation. We also compare two methods for using this feedback: supervised fine-tuning and direct preference optimization. Finally, we introduce SummLlama3-8b, a model that outperforms the nearly 10x larger Llama3-70b-instruct in generating human-preferred summaries, demonstrating that smaller models can achieve superior performance with appropriate training. The full dataset and SummLlama3-8B model are available at https://huggingface.co/datasets/DISLab/FeedSum and https://huggingface.co/DISLab/SummLlama3-8B.

Training automatic summary fact verifiers often faces the challenge of a lack of human-labeled data. In this paper, we explore alternative way of leveraging Large Language Model (LLM) generated feedback to address the inherent limitation of using human-labeled data. We introduce FineSumFact, a large-scale dataset containing fine-grained factual feedback on summaries. We employ 10 distinct LLMs for diverse summary generation and Llama-3-70B-Instruct for feedback. We utilize this dataset to fine-tune the lightweight open-source model Llama-3-8B-Instruct, optimizing resource efficiency while maintaining high performance. Our experimental results reveal that the model trained on extensive LLM-generated datasets surpasses that trained on smaller human-annotated datasets when evaluated using human-generated test sets. Fine-tuning fact verification models with LLM feedback can be more effective and cost-efficient than using human feedback. The dataset is available at https://github.com/DISL-Lab/FineSumFact.