We present CAGSR-vLLM-MTC, an extension of our Self-Supervised Cross-Attention-Guided Reinforcement (CAGSR) framework, now implemented on the high-performance vLLM runtime, to address both multi-turn dialogue and chain-of-thought reasoning. Building upon our original single-turn approach, we first instrumented vLLM's C++/CUDA kernels to asynchronously capture per-layer, per-head cross-attention weights during generation. We then generalized our self-supervised reward function to accumulate attention signals over entire conversation histories and intermediate chain-of-thought steps. We discuss practical trade-offs, including an entropy-based clamping mechanism to prevent attention collapse on early context, and outline future directions for multi-party dialogues and hierarchical reasoning.

Large language models (LLMs) possess a remarkable ability to perform in-context learning (ICL), which enables them to handle multiple downstream tasks simultaneously without requiring task-specific fine-tuning. Recent studies have shown that even moderately sized LLMs, such as Mistral 7B, Gemma 7B and Llama-3 8B, can achieve ICL through few-shot in-context fine-tuning of all tasks at once. However, this approach still lags behind dedicated fine-tuning, where a separate model is trained for each individual task. In this paper, we propose a novel approach, Many-Shot In-Context Fine-tuning (ManyICL), which significantly narrows this performance gap by extending the principles of ICL to a many-shot setting. To unlock the full potential of ManyICL and address the inherent inefficiency of processing long sequences with numerous in-context examples, we propose a novel training objective. Instead of solely predicting the final answer, our approach treats every answer within the context as a supervised training target. This effectively shifts the role of many-shot examples from prompts to targets for autoregressive learning. Through extensive experiments on diverse downstream tasks, including classification, summarization, question answering, natural language inference, and math, we demonstrate that ManyICL substantially outperforms zero/few-shot fine-tuning and approaches the performance of dedicated fine-tuning. Furthermore, ManyICL significantly mitigates catastrophic forgetting issues observed in zero/few-shot fine-tuning. The code will be made publicly available upon publication.

Recent preference learning frameworks for large language models (LLMs) simplify human preferences with binary pairwise comparisons and scalar rewards. This simplification could make LLMs' responses biased to mostly preferred features, and would be exacerbated during the iterations of online preference learning steps. To address these challenges, we propose a novel framework coined PFP (Preference Feature Preservation). The key idea of PFP is maintaining the distribution of human preference features and utilizing such rich signals throughout the online preference learning process. Specifically, PFP first extract preference features from offline pairwise human preference data and trains a feature classifier. Then, using trained classifier and the distribution preserving optimization, PFP maps appropriate preference features for a new input instruction during online learning. Lastly, PFP trains LLM using the existing preference learning method, by incorporating the preference feature into system prompts and enabling LLM to explicitly handle various human preferences. Our experiments demonstrate that PFP successfully mitigates the bias in preference features during online learning, and hence achieves superior performance compared to previous preference learning methods on standard benchmarks to evaluate LLM alignment.

The expanding Lean 4 ecosystem poses challenges for navigating its vast libraries. This paper introduces LeanExplore, a search engine for Lean 4 declarations. LeanExplore enables users to semantically search for statements, both formally and informally, across select Lean 4 packages (including Batteries, Init, Lean, Mathlib, PhysLean, and Std). This search capability is powered by a hybrid ranking strategy, integrating scores from a multi-source semantic embedding model (capturing conceptual meaning from formal Lean code, docstrings, AI-generated informal translations, and declaration titles), BM25+ for keyword-based lexical relevance, and a PageRank-based score reflecting declaration importance and interconnectedness. The search engine is accessible via a dedicated website (https://www.leanexplore.com/) and a Python API (https://github.com/justincasher/lean-explore). Furthermore, the database can be downloaded, allowing users to self-host the service. LeanExplore integrates easily with LLMs via the model context protocol (MCP), enabling users to chat with an AI assistant about Lean declarations or utilize the search engine for building theorem-proving agents. This work details LeanExplore's architecture, data processing, functionalities, and its potential to enhance Lean 4 workflows and AI-driven mathematical research

--We introduce PRISM (Predictive Reasoning in Sequential Medicine), a transformer-based architecture designed to model the sequential progression of clinical decision-making processes. Unlike traditional approaches that rely on isolated diagnostic classification, PRISM frames clinical trajectories as tokenized sequences of events -- including diagnostic tests, laboratory results, and diagnoses -- and learns to predict the most probable next steps in the patient diagnostic journey. Leveraging a large custom clinical vocabulary and an autoregressive training objective, PRISM demonstrates the ability to capture complex dependencies across longitudinal patient timelines. Experimental results show substantial improvements over random baselines in next-token prediction tasks, with generated sequences reflecting realistic diagnostic pathways, laboratory result progressions, and clinician ordering behaviors. These findings highlight the feasibility of applying generative language modeling techniques to structured medical event data, enabling applications in clinical decision support, simulation, and education. PRISM establishes a foundation for future advancements in sequence-based healthcare modeling, bridging the gap between machine learning architectures and real-world diagnostic reasoning. Accurate and timely clinical decision-making is fundamental to high-quality patient care.

Grammar-based test case generation has proven effective for competitive programming problems, but generating valid and general grammars from natural language specifications remains a key challenge, especially under limited supervision. Context-Free Grammars with Counters (CCFGs) have recently been introduced as a formalism to represent such specifications with logical constraints by storing and reusing counter values during derivation. In this work, we explore the use of open-source large language models (LLMs) to induce CCFGs from specifications using a small number of labeled examples and verifiable reward-guided reinforcement learning. Our approach first fine-tunes an open-source LLM to perform specification-to-grammar translation, and further applies Group Relative Policy Optimization (GRPO) to enhance grammar validity and generality. We also examine the effectiveness of iterative feedback for open and closed-source LLMs in correcting syntactic and semantic errors in generated grammars. Experimental results show that our approach SAGE achieves stronger generalization and outperforms 17 open and closed-source LLMs in both grammar quality and test effectiveness, improving over the state-of-the-art by 15.92%p in grammar validity and 12.34%p in test effectiveness. We provide our implementation and dataset at the following anonymous repository:https://anonymous.4open.science/r/SAGE-5714

The rapid advancement of Multimodal Large Language Models (MLLMs) has ignited discussions regarding their potential to surpass human performance in multimodal tasks. In response, we introduce MANBench (Multimodal Ability Norms Benchmark), a bilingual benchmark (English and Chinese) comprising 1,314 questions across nine tasks, spanning knowledge-based and non-knowledge-based domains. MANBench emphasizes intuitive reasoning, seamless cross-modal integration, and real-world complexity, providing a rigorous evaluation framework. Through extensive human experiments involving diverse participants, we compared human performance against state-of-the-art MLLMs. The results indicate that while MLLMs excel in tasks like Knowledge and Text-Image Understanding, they struggle with deeper cross-modal reasoning tasks such as Transmorphic Understanding, Image Consistency, and Multi-image Understanding. Moreover, both humans and MLLMs face challenges in highly complex tasks like Puzzles and Spatial Imagination. MANBench highlights the strengths and limitations of MLLMs, revealing that even advanced models fall short of achieving human-level performance across many domains. We hope MANBench will inspire efforts to bridge the gap between MLLMs and human multimodal capabilities. The code and dataset are available at https://github.com/micdz/MANBench.

Multimodal Retrieval-Augmented Generation (RAG) systems have become essential in knowledge-intensive and open-domain tasks. As retrieval complexity increases, ensuring the robustness of these systems is critical. However, current RAG models are highly sensitive to the order in which evidence is presented, often resulting in unstable performance and biased reasoning, particularly as the number of retrieved items or modality diversity grows. This raises a central question: How does the position of retrieved evidence affect multimodal RAG performance? To answer this, we present the first comprehensive study of position bias in multimodal RAG systems. Through controlled experiments across text-only, image-only, and mixed-modality tasks, we observe a consistent U-shaped accuracy curve with respect to evidence position. To quantify this bias, we introduce the Position Sensitivity Index ($PSI_p$) and develop a visualization framework to trace attention allocation patterns across decoder layers. Our results reveal that multimodal interactions intensify position bias compared to unimodal settings, and that this bias increases logarithmically with retrieval range. These findings offer both theoretical and empirical foundations for position-aware analysis in RAG, highlighting the need for evidence reordering or debiasing strategies to build more reliable and equitable generation systems.

Large Language Model (LLM)-based coding agents have shown promising results on coding benchmarks, but their effectiveness on systems code remains underexplored. Due to the size and complexities of systems code, making changes to a systems codebase is a daunting task, even for humans. It requires researching about many pieces of context, derived from the large codebase and its massive commit history, before making changes. Inspired by the recent progress on deep research agents, we design the first deep research agent for code, called Code Researcher, and apply it to the problem of generating patches for mitigating crashes reported in systems code. Code Researcher performs multi-step reasoning about semantics, patterns, and commit history of code to gather sufficient context. The context is stored in a structured memory which is used for synthesizing a patch. We evaluate Code Researcher on kBenchSyz, a benchmark of Linux kernel crashes, and show that it significantly outperforms strong baselines, achieving a crash-resolution rate of 58%, compared to 37.5% by SWE-agent. On an average, Code Researcher explores 10 files in each trajectory whereas SWE-agent explores only 1.33 files, highlighting Code Researcher's ability to deeply explore the codebase. Through another experiment on an open-source multimedia software, we show the generalizability of Code Researcher. Our experiments highlight the importance of global context gathering and multi-faceted reasoning for large codebases.

Large language models (LLMs) have become integral to modern software development, producing vast amounts of AI-generated source code. While these models boost programming productivity, their misuse introduces critical risks, including code plagiarism, license violations, and the propagation of insecure programs. As a result, robust detection of AI-generated code is essential. To support the development of such detectors, a comprehensive benchmark that reflects real-world conditions is crucial. However, existing benchmarks fall short -- most cover only a limited set of programming languages and rely on less capable generative models. In this paper, we present CodeMirage, a comprehensive benchmark that addresses these limitations through three major advancements: (1) it spans ten widely used programming languages, (2) includes both original and paraphrased code samples, and (3) incorporates outputs from ten state-of-the-art production-level LLMs, including both reasoning and non-reasoning models from six major providers. Using CodeMirage, we evaluate ten representative detectors across four methodological paradigms under four realistic evaluation configurations, reporting results using three complementary metrics. Our analysis reveals nine key findings that uncover the strengths and weaknesses of current detectors, and identify critical challenges for future work. We believe CodeMirage offers a rigorous and practical testbed to advance the development of robust and generalizable AI-generated code detectors.