Document chunking is a crucial component of Retrieval-Augmented Generation (RAG), as it directly affects the retrieval of relevant and precise context. Conventional fixed-length and recursive splitters often produce arbitrary, incoherent segments that fail to preserve semantic structure. Although semantic chunking has gained traction, its influence on generation quality remains underexplored. This paper introduces two efficient semantic chunking methods, Projected Similarity Chunking (PSC) and Metric Fusion Chunking (MFC), trained on PubMed data using three different embedding models. We further present an evaluation framework that measures the effect of chunking on both retrieval and generation by augmenting PubMedQA with full-text PubMed Central articles. Our results show substantial retrieval improvements ( 24x with PSC) in MRR and higher Hits@k on PubMedQA. We provide a comprehensive analysis, including statistical significance and response-time comparisons with common chunking libraries. Despite being trained on a single domain, PSC and MFC also generalize well, achieving strong out-of-domain generation performance across multiple datasets. Overall, our findings confirm that our semantic chunkers, especially PSC, consistently deliver superior performance.

We study timestamped question answering over educational lecture videos under a single-GPU latency/memory budget. Given a natural-language query, the system retrieves relevant timestamped segments and synthesizes a grounded answer. We present CourseTimeQA (52.3 h, 902 queries across six courses) and a lightweight, latency-constrained cross-modal retriever (CrossFusion-RAG) that combines frozen encoders, a learned 512->768 vision projection, shallow query-agnostic cross-attention over ASR and frames with a temporal-consistency regularizer, and a small cross-attentive reranker. On CourseTimeQA, CrossFusion-RAG improves nDCG@10 by 0.10 and MRR by 0.08 over a strong BLIP-2 retriever while achieving approximately 1.55 s median end-to-end latency on a single A100. Closest comparators (zero-shot CLIP multi-frame pooling; CLIP + cross-encoder reranker + MMR; learned late-fusion gating; text-only hybrid with cross-encoder reranking and its MMR variant; caption-augmented text retrieval; non-learned temporal smoothing) are evaluated under matched hardware and indexing. We report robustness across ASR noise (WER quartiles), diagnostics for temporal localization, and full training/tuning details to support reproducible comparison.

Are frontier AI systems becoming more capable? Certainly. Yet such progress is not an unalloyed blessing but rather a Trojan horse: behind their performance leaps lie more insidious and destructive safety risks, namely deception. Unlike hallucination, which arises from insufficient capability and leads to mistakes, deception represents a deeper threat in which models deliberately mislead users through complex reasoning and insincere responses. As system capabilities advance, deceptive behaviours have spread from textual to multimodal settings, amplifying their potential harm. First and foremost, how can we monitor these covert multimodal deceptive behaviors? Nevertheless, current research remains almost entirely confined to text, leaving the deceptive risks of multimodal large language models unexplored. In this work, we systematically reveal and quantify multimodal deception risks, introducing MM-DeceptionBench, the first benchmark explicitly designed to evaluate multimodal deception. Covering six categories of deception, MM-DeceptionBench characterizes how models strategically manipulate and mislead through combined visual and textual modalities. On the other hand, multimodal deception evaluation is almost a blind spot in existing methods. Its stealth, compounded by visual-semantic ambiguity and the complexity of cross-modal reasoning, renders action monitoring and chain-of-thought monitoring largely ineffective. To tackle this challenge, we propose debate with images, a novel multi-agent debate monitor framework. By compelling models to ground their claims in visual evidence, this method substantially improves the detectability of deceptive strategies. Experiments show that it consistently increases agreement with human judgements across all tested models, boosting Cohen's kappa by 1.5x and accuracy by 1.25x on GPT-4o.

The LLM field has spent a year perfecting RL for tasks machines already excel at, math, code, and deterministic reasoning, while completely sidestepping the domain that actually defines human intelligence: subjective, emotionally grounded, personality sensitive conversation. This space has often been regarded as inherently subjective and challenging to formalize, making it appear unsuitable for conventional RL pipelines. We show that it is not only possible and it is a solvable and transformative RL problem. We propose the first framework that infers user personality on the fly and optimizes model behavior toward personalized conversational preferences. Contrary to the widespread belief that RL collapses in non-verifiable settings, our method produces consistent, robust, and dramatic improvements in humanlike interaction quality. We also introduce the first dynamic emotional intelligence evaluation suite to quantify these gains. Our model, which is introduced as Echo-N1, behaves far above its base version and outperforming the proprietary Doubao 1.5 Character. This work establishes a new frontier for RL: optimizing models for the deeply subjective, deeply human dimensions of conversation.

Large language models (LLMs) are increasingly deployed as conversational tutors in STEM education, yet most systems still rely on a single LLM with a static retrieval-augmented generation (RAG) pipeline over course materials. This design struggles in complex domains such as digital signal processing (DSP), where tutors must maintain coherent long-term student models, manage heterogeneous knowledge bases, and adapt teaching strategies over extended interactions. We argue that retrieval, memory, and control should be treated as a coupled cognitive evolution process. We instantiate this view in CogEvo-Edu, a hierarchical educational multi-agent system comprising a Cognitive Perception Layer (CPL), a Knowledge Evolution Layer (KEL), and a Meta-Control Layer (MCL). CPL maintains dual memories and performs confidence-weighted consolidation to build structured, self-correcting student profiles under limited context. KEL assigns each knowledge chunk a spatiotemporal value that drives activation, semantic compression, and forgetting. MCL formulates tutoring as hierarchical sequential decision making, orchestrating specialized agents and jointly adapting CPL/KEL hyperparameters via a dual inner--outer loop. To evaluate CogEvo-Edu, we construct DSP-EduBench, a vertical benchmark for DSP tutoring with heterogeneous resources, simulated student profiles, and long-horizon interaction scripts. Using a three-model LLM-as-a-Judge ensemble, CogEvo-Edu raises the overall score from 5.32 to 9.23 and improves all six indicators over static RAG, simple memory, and a single-agent variant, demonstrating the value of jointly evolving student profiles, knowledge bases, and teaching policies.

Temporal reasoning over evolving semi-structured tables poses a challenge to current QA systems. We propose a SQL-based approach that involves (1) generating a 3NF schema from Wikipedia infoboxes, (2) generating SQL queries, and (3) query execution. Our central finding challenges model scaling assumptions: the quality of schema design has a greater impact on QA precision than model capacity. We establish three evidence-based principles: normalization that preserves context, semantic naming that reduces ambiguity, and consistent temporal anchoring. Our best configuration (Gemini 2.5 Flash schema + Gemini-2.0-Flash queries) achieves 80.39 EM, a 16.8\% improvement over the baseline (68.89 EM).

Bug reports are often unstructured and verbose, making it challenging for developers to efficiently comprehend software issues. Existing summarization approaches typically rely on surface-level textual cues, resulting in incomplete or redundant summaries, and they frequently ignore associated code snippets, which are essential for accurate defect diagnosis. To address these limitations, we propose a progressive code-integration framework for LLM-based abstractive bug report summarization. Our approach incrementally incorporates long code snippets alongside textual content, overcoming standard LLM context window constraints and producing semantically rich summaries. Evaluated on four benchmark datasets using eight LLMs, our pipeline outperforms extractive baselines by 7.5%-58.2% and achieves performance comparable to state-of-the-art abstractive methods, highlighting the benefits of jointly leveraging textual and code information for enhanced bug comprehension.

Context-based question answering (CBQA) models provide more accurate and relevant answers by considering the contextual information. They effectively extract specific information given a context, making them functional in various applications involving user support, information retrieval, and educational platforms. In this manuscript, we benchmarked the performance of 47 CBQA models from Hugging Face on eight different datasets. This study aims to identify the best-performing model across diverse datasets without additional fine-tuning. It is valuable for practical applications where the need to retrain models for specific datasets is minimized, streamlining the implementation of these models in various contexts. The best-performing models were trained on the SQuAD v2 or SQuAD v1 datasets. The best-performing model was ahotrod/electra_large_discriminator_squad2_512, which yielded 43\% accuracy across all datasets. We observed that the computation time of all models depends on the context length and the model size. The model's performance usually decreases with an increase in the answer length. Moreover, the model's performance depends on the context complexity. We also used the Genetic algorithm to improve the overall accuracy by integrating responses from other models. ahotrod/electra_large_discriminator_squad2_512 generated the best results for bioasq10b-factoid (65.92\%), biomedical\_cpgQA (96.45\%), QuAC (11.13\%), and Question Answer Dataset (41.6\%). Bert-large-uncased-whole-word-masking-finetuned-squad achieved an accuracy of 82\% on the IELTS dataset.

Virtual cell modeling aims to predict cellular responses to perturbations. Existing virtual cell models rely heavily on large-scale single-cell datasets, learning explicit mappings between gene expression and perturbations. Although recent models attempt to incorporate multi-source biological information, their generalization remains constrained by data quality, coverage, and batch effects. More critically, these models often function as black boxes, offering predictions without interpretability or consistency with biological principles, which undermines their credibility in scientific research. To address these challenges, we present VCWorld, a cell-level white-box simulator that integrates structured biological knowledge with the iterative reasoning capabilities of large language models to instantiate a biological world model. VCWorld operates in a data-efficient manner to reproduce perturbation-induced signaling cascades and generates interpretable, stepwise predictions alongside explicit mechanistic hypotheses. In drug perturbation benchmarks, VCWorld achieves state-of-the-art predictive performance, and the inferred mechanistic pathways are consistent with publicly available biological evidence.

Multimodal Large Language Models (MLLMs) have emerged as powerful tools for chart comprehension. However, they heavily rely on extracted content via OCR, which leads to numerical hallucinations when chart textual annotations are sparse. While existing methods focus on scaling instructions, they fail to address the fundamental challenge, i.e., reasoning with visual perception. In this paper, we identify a critical observation: MLLMs exhibit weak grounding in chart elements and proportional relationships, as evidenced by their inability to localize key positions to match their reasoning. To bridge this gap, we propose PointCoT, which integrates reflective interaction into chain-of-thought reasoning in charts. By prompting MLLMs to generate bounding boxes and re-render charts based on location annotations, we establish connections between textual reasoning steps and visual grounding regions. We further introduce an automated pipeline to construct ChartPoint-SFT-62k, a dataset featuring 19.2K high-quality chart samples with step-by-step CoT, bounding box, and re-rendered visualizations. Leveraging this data, we develop two instruction-tuned models, ChartPointQ2 and ChartPointQ2.5, which outperform state-of-the-art across several chart benchmarks, e.g., +5.04\% on ChartBench.