Retrieval-augmented language models improve knowledge-intensive question answering, but they remain brittle on multi-step questions, where an early retrieval error can steer subsequent reasoning toward the wrong entity or relation. Existing methods can often detect that the current answer or evidence is unreliable, yet they rarely turn this diagnosis into an explicit retrieval correction. We propose Verification-Guided Retrieval Revision (VRR), a framework that uses verification as a control signal for multi-step retrieval. VRR couples a Generator that proposes search and answers, a Judge that verifies answer correctness and evidential support while emitting structured failure diagnoses, and a Retrieval Context Manager (RCM) that converts these diagnoses into retriever-facing updates such as clue selection, query rewriting, and state-aware evidence ranking. To make this loop learnable, we train VRR in three stages: iterative cold-start training to stabilize search, judge-shaped policy optimization to improve long-horizon reasoning, and diagnosis-conditioned RCM training for retrieval adaptation. Across seven open-domain QA benchmarks with a fixed retriever and corpus, training on HotpotQA and transferring to out-of-domain datasets, VRR surpasses the strongest prior baseline by 4.2 EM on average under budget-aligned evaluation, with larger gains on multi-hop benchmarks (+5.4).

Existing cross-modal image-text retrieval models often retrieve samples with inconsistent details. To evaluate fine-grained discriminability, we introduce MSCOCO-CCD and Flickr30k-CCD, with three key features: (1) a two-level image content taxonomy for contrastive sample generation and fine-grained evaluation; (2) annotation of numerous contrastive samples, where each sample differs from the anchor by a controlled contrastive difference (CCD), with the specific type of difference labeled; (3) a fine-grained contrastive discrimination metric to assess the ability to distinguish fine-grained nuances. Extensive experiments demonstrate that contrastive samples can significantly degrade retrieval performance. Furthermore, fine-grained evaluation reveals that current models still struggle to effectively produce discriminative representations on certain feature types, such as entity emotion and scene attribute. Our datasets and related codes will be publicly released.

Large language models (LLMs) possess extensive world knowledge and have demonstrated strong potential for misinformation detection. A growing line of recent work prompts LLMs to produce structured reasoning before issuing a veracity judgment. However, we contend that such structured outputs do not spontaneously support the veracity decision. The emitted scores can satisfy the output schema while contributing little to veracity discrimination. Since annotating full reasoning paths is prohibitively costly for veracity tasks, we propose a Generation-Score Consistency framework that supervises through structured dimension scores and steers the LLM toward functional reasoning. Specifically, we train a compact score-to-label verifier on a small set of expert-annotated reasoning data, freeze it, and during supervised fine-tuning use a trainable score head to align the score-token hidden states of the LLM with the human-calibrated mapping. Introducing structured reasoning on its own costs accuracy relative to label-only supervision, while the consistency term recovers that cost and keeps the emitted scores coupled to the veracity decision. Controlled experiments on multimodal misinformation benchmarks show that the consistency term strengthens both veracity prediction and the discriminability of the emitted scores. Code and data are available at https://github.com/yun-nlp/Reasoning-Scores-for-Misinformation-Detection-EMNLP2026.

Text-to-image models can render recognizable cultural symbols, but cultural activity generation requires coordinated control over participants, actions, interactions, scenes, and culture-specific constraints. We propose CARE, an evidence-grounded visual checklist framework for controllable cultural activity image generation. Given an activity prompt, CARE builds source-traceable cultural evidence, converts it into an optimized prompt and fine-grained checklist, verifies generated images with concrete VLM-based visual checks, and uses failed checks for targeted prompt repair. We introduce CARE-1360, a dataset of 1,360 cultural activity prompts across eight countries, and evaluate generated images with a unified human protocol. Experiments show that CARE consistently improves over direct generation and retrieval-based prompt refinement baselines across cultural and visual quality dimensions. Results with both ZImage and Stable Diffusion XL suggest that evidence-grounded visual checklists provide an interpretable control mechanism that transfers across the tested generation backends.

Temporal knowledge graph forecasting predicts future facts from historical snapshots, and its accuracy hinges on how well a model exploits the historical evidence available to each query. Existing methods route every query through a single reasoning pipeline, yet queries with the same predicate form can rest on very different histories and call for different inductive biases. We observe that query histories fall into three regimes: sparse histories with few prior events, recurrent histories dominated by stable repetition, and ambiguous histories with many competing candidates. Building on this observation, we propose QTI-MoE, a Query-Type-aware Inductive-bias Mixture-of-Experts framework. QTI-MoE characterises each query with a soft type label derived from its historical prefix and forwards it to one of three experts whose memory pathways carry complementary biases: recency-weighted aggregation for sparse queries, count-based repetition for recurrent queries, and diversity-normalised scoring for ambiguous queries. QTI-MoE achieves state-of-the-art performance on four standard benchmarks, and a type-stratified analysis confirms that each expert peaks on its target regime.

Large Language Models (LLMs) exhibit strong reasoning capabilities and are widely applied in event forecasting. However, studies have demonstrated that LLMs exhibit human-like cognitive biases, systematic patterns of deviation from rationality in decision-making. To explore the cognitive biases in event forecasting, we introduce CogForecast, a human-curated dataset comprising six topics. Experimental results on three LLMs reveal significant cognitive biases in LLM-based event forecasting methods. To address this issue, we propose MCA, a Multi-Cognition Agentic framework. Specifically, MCA leverages LLMs to act as multi-cognition event participants, performing perspective-taking based on the cognitive patterns of event participants to alleviate the inherent cognitive biases in LLMs and offer diverse analytical perspectives. Then, MCA clusters agents according to their predictions and derives a final answer through a group-level reliability scoring method. Experimental results on a dataset including eight event categories demonstrate the effectiveness of MCA. Using Llama-3.1-70B, MCA achieves an accuracy of 82.3% (79.5% for the human crowd). Additionally, we demonstrate that MCA can alleviate the cognitive biases in LLMs and investigate three influencing factors.

Prompt-based learning, which exploits knowledge from pre-trained language models by providing textual prompts and designing appropriate answer-category mapping methods, has achieved impressive successes on few-shot text classification and natural language inference (NLI). Because of the diverse linguistic expression, there exist many answer tokens for the same category. However, both manual answer design and automatic answer search constrain answer space and therefore hardly achieve ideal performance. To address this issue, we propose an answer space clustered prompting model (ASCM) together with a synonym initialization method (SI) which automatically categorizes all answer tokens in a semantic-clustered embedding space. We also propose a stable semi-supervised method named stair learning (SL) that orderly distills knowledge from better models to weaker models. Extensive experiments demonstrate that our ASCM+SL significantly outperforms existing state-of-the-art techniques in few-shot settings.