Radiography Report Generation (RRG) has gained significant attention in medical image analysis as a promising tool for alleviating the growing workload of radiologists. However, despite numerous advancements, existing methods have yet to achieve the quality required for deployment in real-world clinical settings. Meanwhile, large Visual Language Models (VLMs) have demonstrated remarkable progress in the general domain by adopting training strategies originally designed for Large Language Models (LLMs), such as alignment techniques. In this paper, we introduce a model-agnostic framework to enhance RRG accuracy using Direct Preference Optimization (DPO). Our approach leverages random contrastive sampling to construct training pairs, eliminating the need for reward models or human preference annotations. Experiments on supplementing three state-of-the-art models with our Random DPO show that our method improves clinical performance metrics by up to 5%, without requiring any additional training data.

White box monitors that analyze model internals offer promising advantages for detecting potentially harmful behaviors in large language models, including lower computational costs and integration into layered defense systems.However, training and evaluating these monitors requires response exemplars that exhibit the target behaviors, typically elicited through prompting or fine-tuning. This presents a challenge when the information used to elicit behaviors inevitably leaks into the data that monitors ingest, inflating their effectiveness. We present a systematic framework for evaluating a monitor's performance in terms of its ability to detect genuine model behavior rather than superficial elicitation artifacts. Furthermore, we propose three novel strategies to evaluate the monitor: content filtering (removing deception-related text from inputs), score filtering (aggregating only over task-relevant tokens), and prompt distilled fine-tuned model organisms (models trained to exhibit deceptive behavior without explicit prompting). Using deception detection as a representative case study, we identify two forms of leakage that inflate monitor performance: elicitation leakage from prompts that explicitly request harmful behavior, and reasoning leakage from models that verbalize their deceptive actions. Through experiments on multiple deception benchmarks, we apply our proposed mitigation strategies and measure performance retention. Our evaluation of the monitors reveal three crucial findings: (1) Content filtering is a good mitigation strategy that allows for a smooth removal of elicitation signal and can decrease probe AUROC by 30\% (2) Score filtering was found to reduce AUROC by 15\% but is not as straightforward to attribute to (3) A finetuned model organism improves monitor evaluations but reduces their performance by upto 40\%, even when re-trained.

Retrieval-augmented generation (RAG) has become a dominant paradigm for mitigating knowledge hallucination and staleness in large language models (LLMs) while preserving data security. By retrieving relevant evidence from private, domain-specific corpora and injecting it into carefully engineered prompts, RAG delivers trustworthy responses without the prohibitive cost of fine-tuning. Traditional retrieval-augmented generation (RAG) systems are text-only and often rely on a single storage backend, most commonly a vector database. In practice, this monolithic design suffers from unavoidable trade-offs: vector search captures semantic similarity yet loses global context; knowledge graphs excel at relational precision but struggle with recall; full-text indexes are fast and exact yet semantically blind; and relational engines such as MySQL provide strong transactional guarantees but no semantic understanding. We argue that these heterogeneous retrieval paradigms are complementary, and propose a principled fusion scheme to orchestrate them synergistically, mitigating the weaknesses of any single modality. In this work we introduce HetaRAG, a hybrid, deep-retrieval augmented generation framework that orchestrates cross-modal evidence from heterogeneous data stores. We plan to design a system that unifies vector indices, knowledge graphs, full-text engines, and structured databases into a single retrieval plane, dynamically routing and fusing evidence to maximize recall, precision, and contextual fidelity. To achieve this design goal, we carried out preliminary explorations and constructed an initial RAG pipeline; this technical report provides a brief overview. The partial code is available at https://github.com/KnowledgeXLab/HetaRAG.

Retrieval-Augmented Generation (RAG) has emerged as the standard paradigm for answering questions on enterprise data. Traditionally, RAG has centered on text-based semantic search and re-ranking. However, this approach falls short when dealing with questions beyond data summarization or non-text data. This has led to various attempts to supplement RAG to bridge the gap between RAG, the implementation paradigm, and the question answering problem that enterprise users expect it to solve. Given that contemporary RAG is a collection of techniques rather than a defined implementation, discussion of RAG and related question-answering systems benefits from a problem-oriented understanding. We propose a new classification framework (L1-L5) to categorize systems based on data modalities and task complexity of the underlying question answering problems: L1 (Surface Knowledge of Unstructured Data) through L4 (Reflective and Reasoned Knowledge) and the aspirational L5 (General Intelligence). We also introduce benchmarks aligned with these levels and evaluate four state-of-the-art platforms: LangChain, Azure AI Search, OpenAI, and Corvic AI. Our experiments highlight the value of multi-space retrieval and dynamic orchestration for enabling L1-L4 capabilities. We empirically validate our findings using diverse datasets indicative of enterprise use cases.

Generative recommendation has emerged as a promising paradigm, demonstrating remarkable results in both academic benchmarks and industrial applications. However, existing systems predominantly focus on unifying retrieval and ranking while neglecting the integration of search and recommendation (S&R) tasks. What makes search and recommendation different is how queries are formed: search uses explicit user requests, while recommendation relies on implicit user interests. As for retrieval versus ranking, the distinction comes down to whether the queries are the target items themselves. Recognizing the query as central element, we propose IntSR, an integrated generative framework for S&R. It also addresses the increased computational complexity associated with integrated S&R behaviors and the erroneous pattern learning introduced by a dynamically changing corpus. IntSR has been successfully deployed across various scenarios in Amap, leading to substantial improvements in digital asset's GMV(+9.34%), Search and recommendation (S&R) services are now commonly provided by online platforms, such as Y ouTube and Amazon. These two tasks operate on shared users and items, creating a natural foundation for the joint modeling and application of S&R. A unified S&R model can better capture user preferences and enhance the effectiveness of both tasks, while also reducing engineering overhead (the left side of Figure 1).

Recent progress in large reasoning models for challenging mathematical reasoning has been driven by reinforcement learning (RL). Incorporating long chain-of-thought (CoT) data during mid-training has also been shown to substantially improve reasoning depth. However, current approaches often utilize CoT data indiscriminately, leaving open the critical question of which data types most effectively enhance model reasoning capabilities. In this paper, we define the foundation model's reasoning potential for the first time as the inverse of the number of independent attempts required to correctly answer the question, which is strongly correlated with the final model performance. We then propose utilizing diverse data enriched with high-value reasoning patterns to expand the reasoning potential. Specifically, we abstract atomic reasoning patterns from CoT sequences, characterized by commonality and inductive capabilities, and use them to construct a core reference set enriched with valuable reasoning patterns. Furthermore, we propose a dual-granularity algorithm involving chains of reasoning patterns and token entropy, efficiently selecting high-value CoT data (CoTP) from the data pool that aligns with the core set, thereby training models to master reasoning effectively. Only 10B-token CoTP data enables the 85A6B Mixture-of-Experts (MoE) model to improve by 9.58% on the challenging AIME 2024 and 2025, and to raise the upper bound of downstream RL performance by 7.81%.

The adoption of Large Language Models (LLMs) as automated evaluators (LLM-as-a-judge) has revealed critical inconsistencies in current evaluation frameworks. We identify two fundamental types of inconsistencies: (1) Score-Comparison Inconsistency, where lower-rated responses outperform higher-scored ones in pairwise comparisons, and (2) Pairwise Transitivity Inconsistency, manifested through circular preference chains (A>B>C>A) and equivalence contradictions (A=B=C\neq A). We argue that these issues come from information loss in discrete rating systems and ambiguous tie judgments during pairwise evaluation. We propose TrustJudge, a probabilistic framework that addresses these limitations through two key innovations: 1) distribution-sensitive scoring that computes continuous expectations from discrete rating probabilities, preserving information entropy for more precise scoring, and 2) likelihood-aware aggregation that resolves transitivity violations using bidirectional preference probabilities or perplexity. We also formalize the theoretical limitations of current LLM-as-a-judge frameworks and demonstrate how TrustJudge's components overcome them. When evaluated with Llama-3.1-70B-Instruct as judge using our dataset, TrustJudge reduces Score-Comparison inconsistency by 8.43% (from 23.32% to 14.89%) and Pairwise Transitivity inconsistency by 10.82% (from 15.22% to 4.40%), while maintaining higher evaluation accuracy. Our work provides the first systematic analysis of evaluation framework inconsistencies in LLM-as-a-judge paradigms, offering both theoretical insights and practical solutions for reliable automated assessment. The framework demonstrates consistent improvements across various model architectures and scales, enabling more trustworthy LLM evaluation without requiring additional training or human annotations. The codes can be found at https://github.com/TrustJudge/TrustJudge.

The integration of new modalities enhances the capabilities of multimodal large language models (MLLMs) but also introduces additional vulnerabilities. In particular, simple visual jailbreaking attacks can manipulate open-source MLLMs more readily than sophisticated textual attacks. However, these underdeveloped attacks exhibit extremely limited cross-model transferability, failing to reliably identify vulnerabilities in closed-source MLLMs. In this work, we analyse the loss landscape of these jailbreaking attacks and find that the generated attacks tend to reside in high-sharpness regions, whose effectiveness is highly sensitive to even minor parameter changes during transfer. To further explain the high-sharpness localisations, we analyse their feature representations in both the intermediate layers and the spectral domain, revealing an improper reliance on narrow layer representations and semantically poor frequency components. Building on this, we propose a Feature Over-Reliance CorrEction (FORCE) method, which guides the attack to explore broader feasible regions across layer features and rescales the influence of frequency features according to their semantic content. By eliminating non-generalizable reliance on both layer and spectral features, our method discovers flattened feasible regions for visual jailbreaking attacks, thereby improving cross-model transferability. Extensive experiments demonstrate that our approach effectively facilitates visual red-teaming evaluations against closed-source MLLMs.

Large language models (LLMs) have demonstrated notable potential in medical applications, yet they face substantial challenges in handling complex real-world clinical diagnoses using conventional prompting methods. Current prompt engineering and multi-agent approaches typically optimize isolated inferences, neglecting the accumulation of reusable clinical experience. To address this, this study proposes a novel Multi-Agent Clinical Diagnosis (MACD) framework, which allows LLMs to self-learn clinical knowledge via a multi-agent pipeline that summarizes, refines, and applies diagnostic insights. It mirrors how physicians develop expertise through experience, enabling more focused and accurate diagnosis on key disease-specific cues. We further extend it to a MACD-human collaborative workflow, where multiple LLM-based diagnostician agents engage in iterative consultations, supported by an evaluator agent and human oversight for cases where agreement is not reached. Evaluated on 4,390 real-world patient cases across seven diseases using diverse open-source LLMs (Llama-3.1 8B/70B, DeepSeek-R1-Distill-Llama 70B), MACD significantly improves primary diagnostic accuracy, outperforming established clinical guidelines with gains up to 22.3% (MACD). In direct comparison with physician-only diagnosis under the same evaluation protocol, MACD achieves comparable or superior performance, with improvements up to 16%. Furthermore, the MACD-human workflow yields an 18.6% improvement over physician-only diagnosis, demonstrating the synergistic potential of human-AI collaboration. Notably, the self-learned clinical knowledge exhibits strong cross-model stability, transferability across LLMs, and capacity for model-specific personalization.This work thus presents a scalable self-learning paradigm that bridges the gap between the intrinsic knowledge of LLMs.

The scaling of Large Language Models (LLMs) has exposed a critical gap between their performance on static benchmarks and their fragility in dynamic, information-rich environments. While models excel at isolated tasks, the computational limits that govern their reasoning under cognitive load remain poorly understood. In this work, we introduce a formal theory of computational cognitive load, positing that extraneous, task-irrelevant information (Context Saturation) and interference from task-switching (Attentional Residue) are key mechanisms that degrade performance. We designed the Interleaved Cognitive Evaluation (ICE), a deconfounded benchmark to systematically manipulate these load factors on challenging multi-hop reasoning tasks. A comprehensive study (N = 10 replications per item across 200 questions) revealed significant performance variations across five instruction-tuned models. Smaller open-source architectures (Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.2) exhibited baseline brittleness, achieving 0% accuracy (SEM = 0.0) across all conditions, including clean controls, on this high-intrinsic-load task. In contrast, Gemini-2.0-Flash-001 showed partial resilience, achieving 85% accuracy in control conditions, with a statistically significant degradation under context saturation ($β= -0.003$ per % load, $p < 0.001$). These findings provide preliminary evidence that cognitive load is a key contributor to reasoning failures, supporting theories of hallucination-as-guessing under uncertainty. We conclude that dynamic, cognitive-aware stress testing, as exemplified by the ICE benchmark, is essential for evaluating the true resilience and safety of advanced AI systems.