Language model (LM) assistants are increasingly used in applications such as brainstorming and research. Improvements in memory and context size have allowed these models to become more autonomous, which has also resulted in more text accumulation in their context windows without explicit user intervention. This comes with a latent risk: the belief profiles of models -- their understanding of the world as manifested in their responses or actions -- may silently change as context accumulates. This can lead to subtly inconsistent user experiences, or shifts in behavior that deviate from the original alignment of the models. In this paper, we explore how accumulating context by engaging in interactions and processing text -- talking and reading -- can change the beliefs of language models, as manifested in their responses and behaviors. Our results reveal that models' belief profiles are highly malleable: GPT-5 exhibits a 54.7% shift in its stated beliefs after 10 rounds of discussion about moral dilemmas and queries about safety, while Grok 4 shows a 27.2% shift on political issues after reading texts from the opposing position. We also examine models' behavioral changes by designing tasks that require tool use, where each tool selection corresponds to an implicit belief. We find that these changes align with stated belief shifts, suggesting that belief shifts will be reflected in actual behavior in agentic systems. Our analysis exposes the hidden risk of belief shift as models undergo extended sessions of talking or reading, rendering their opinions and actions unreliable.

Compilers, while essential, are notoriously complex systems that demand prohibitively expensive human expertise to develop and maintain. The recent advancements in Large Language Models (LLMs) offer a compelling new paradigm: Neural Compilation, which could potentially simplify compiler development for new architectures and facilitate the discovery of innovative optimization techniques. However, several critical obstacles impede its practical adoption. Firstly, a significant lack of dedicated benchmarks and robust evaluation methodologies hinders objective assessment and tracking of progress in the field. Secondly, systematically enhancing the reliability and performance of LLM-generated assembly remains a critical challenge. Addressing these challenges, this paper introduces NeuComBack, a novel benchmark dataset specifically designed for IR-to-assembly compilation. Leveraging this dataset, we first define a foundational Neural Compilation workflow and conduct a comprehensive evaluation of the capabilities of recent frontier LLMs on Neural Compilation, establishing new performance baselines. We further propose a self-evolving prompt optimization method that enables LLMs to iteratively evolve their internal prompt strategies by extracting insights from prior self-debugging traces, thereby enhancing their neural compilation capabilities. Experiments demonstrate that our method significantly improves both the functional correctness and the performance of LLM-generated assembly code. Compared to baseline prompts, the functional correctness rates improved from 44% to 64% on x86_64 and from 36% to 58% on aarch64, respectively. More significantly, among the 16 correctly generated x86_64 programs using our method, 14 (87.5%) surpassed clang-O3 performance.

Large Language Models (LLMs) have demonstrated remarkable progress in translating natural language to SQL, but a significant semantic gap persists between their general knowledge and domain-specific semantics of databases. Historical translation logs constitute a rich source of this missing in-domain knowledge, where SQL queries inherently encapsulate real-world usage patterns of database schema. Existing methods primarily enhance the reasoning process for individual translations but fail to accumulate in-domain knowledge from past translations. We introduce ORANGE, an online self-evolutionary framework that constructs database-specific knowledge bases by parsing SQL queries from translation logs. By accumulating in-domain knowledge that contains schema and data semantics, ORANGE progressively reduces the semantic gap and enhances the accuracy of subsequent SQL translations. To ensure reliability, we propose a novel nested Chain-of-Thought SQL-to-Text strategy with tuple-semantic tracking, which reduces semantic errors during knowledge generation. Experiments on multiple benchmarks confirm the practicality of ORANGE, demonstrating its effectiveness for real-world Text-to-SQL deployment, particularly in handling complex and domain-specific queries.

Large language models (LLMs) undergo safety alignment after training and tuning, yet recent work shows that safety can be bypassed through jailbreak attacks. While many jailbreaks and defenses exist, their cross-lingual generalization remains underexplored. This paper presents the first systematic multilingual evaluation of jailbreaks and defenses across ten languages -- spanning high-, medium-, and low-resource languages -- using six LLMs on HarmBench and AdvBench. We assess two jailbreak types: logical-expression-based and adversarial-prompt-based. For both types, attack success and defense robustness vary across languages: high-resource languages are safer under standard queries but more vulnerable to adversarial ones. Simple defenses can be effective, but are language- and model-dependent. These findings call for language-aware and cross-lingual safety benchmarks for LLMs.

Sentiment analysis using deep learning and pre-trained language models (PLMs) has gained significant traction due to their ability to capture rich contextual representations. However, existing approaches often underperform in scenarios involving nuanced emotional cues, domain shifts, and imbalanced sentiment distributions. We argue that these limitations stem from inadequate semantic grounding, poor generalization to diverse linguistic patterns, and biases toward dominant sentiment classes. To overcome these challenges, we propose CISEA-MRFE, a novel PLM-based framework integrating Contextual Instruction (CI), Semantic Enhancement Augmentation (SEA), and Multi-Refined Feature Extraction (MRFE). CI injects domain-aware directives to guide sentiment disambiguation; SEA improves robustness through sentiment-consistent paraphrastic augmentation; and MRFE combines a Scale-Adaptive Depthwise Encoder (SADE) for multi-scale feature specialization with an Emotion Evaluator Context Encoder (EECE) for affect-aware sequence modeling. Experimental results on four benchmark datasets demonstrate that CISEA-MRFE consistently outperforms strong baselines, achieving relative improvements in accuracy of up to 4.6% on IMDb, 6.5% on Yelp, 30.3% on Twitter, and 4.1% on Amazon. These results validate the effectiveness and generalization ability of our approach for sentiment classification across varied domains.

The remarkable capabilities of Large Language Models (LLMs) often need to be tailored for specific applications, requiring the integration of new knowledge or the acquisition of new skills. While full fine-tuning is a powerful adaptation method, it is computationally expensive and can lead to a degradation of general reasoning abilities, a phenomenon known as catastrophic forgetting. A range of alternative techniques exists, each with its own trade-offs. In-Context Learning (ICL) is fast but limited by context length, while Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) offer a middle ground by minimizing parameter changes. However, the challenge of catastrophic forgetting persists, raising questions about the best adaptation strategy for a given task. This paper presents a comparative analysis of Supervised Finetuning (SFT), LoRA, and ICL in data-scarce scenarios. We find that LoRA provides the most effective balance, successfully instilling new skills with minimal impact on the base model's general knowledge. In contrast, while SFT excels at skill acquisition, it is highly susceptible to catastrophic forgetting. ICL is effective for incorporating factual knowledge but struggles with complex skills. Our findings offer a practical framework for selecting an LLM adaptation strategy. We highlight the critical distinction between skill acquisition and knowledge integration, clarify the trade-offs between task-specific performance and the preservation of general capabilities.

Federated Learning (FL) is an approach for training a shared Machine Learning (ML) model with distributed training data and multiple participants. FL allows bypassing limitations of the traditional Centralized Machine Learning CL if data cannot be shared or stored centrally due to privacy or technical restrictions -- the participants train the model locally with their training data and do not need to share it among the other participants. This paper investigates the practical implementation and effectiveness of FL in a real-world scenario, specifically focusing on unmanned aerial vehicle (UAV)-based thermal images for common thermal feature detection in urban environments. The distributed nature of the data arises naturally and makes it suitable for FL applications, as images captured in two German cities are available. This application presents unique challenges due to non-identical distribution and feature characteristics of data captured at both locations. The study makes several key contributions by evaluating FL algorithms in real deployment scenarios rather than simulation. We compare several FL approaches with a centralized learning baseline across key performance metrics such as model accuracy, training time, communication overhead, and energy usage. This paper also explores various FL workflows, comparing client-controlled workflows and server-controlled workflows. The findings of this work serve as a valuable reference for understanding the practical application and limitations of the FL methods in segmentation tasks in UAV-based imaging.

In recent years, Neural Operators(NO) have gradually emerged as a popular approach for solving Partial Differential Equations (PDEs). However, their application to large-scale engineering tasks suffers from significant computational overhead. And the fact that current models impose a uniform computational cost while physical fields exhibit vastly different complexities constitutes a fundamental mismatch, which is the root of this inefficiency. For instance, in turbulence flows, intricate vortex regions require deeper network processing compared to stable flows. To address this, we introduce a framework: Skip-Block Routing (SBR), a general framework designed for Transformer-based neural operators, capable of being integrated into their multi-layer architectures. First, SBR uses a routing mechanism to learn the complexity and ranking of tokens, which is then applied during inference. Then, in later layers, it decides how many tokens are passed forward based on this ranking. This way, the model focuses more processing capacity on the tokens that are more complex. Experiments demonstrate that SBR is a general framework that seamlessly integrates into various neural operators. Our method reduces computational cost by approximately 50% in terms of Floating Point Operations (FLOPs), while still delivering up to 2x faster inference without sacrificing accuracy.

A central architectural question for both biological and artificial intelligence is whether judgment relies on specialized modules or a unified, domain-general resource. While the discovery of decodable neural representations for distinct concepts in Large Language Models (LLMs) has suggested a modular architecture, whether these representations are truly independent systems remains an open question. Here we provide evidence for a convergent architecture for evaluative judgment. Across a range of LLMs, we find that diverse evaluative judgments are computed along a dominant dimension, which we term the Valence-Assent Axis (VAA). This axis jointly encodes subjective valence ("what is good") and the model's assent to factual claims ("what is true"). Through direct interventions, we demonstrate this axis drives a critical mechanism, which is identified as the subordination of reasoning: the VAA functions as a control signal that steers the generative process to construct a rationale consistent with its evaluative state, even at the cost of factual accuracy. Our discovery offers a mechanistic account for response bias and hallucination, revealing how an architecture that promotes coherent judgment can systematically undermine faithful reasoning.

As AI systems become increasingly integrated into human lives, endowing them with robust social intelligence has emerged as a critical frontier. A key aspect of this intelligence is discerning truth from deception, a ubiquitous element of human interaction that is conveyed through a complex interplay of verbal language and non-verbal visual cues. However, automatic deception detection in dynamic, multi-party conversations remains a significant challenge. The recent rise of powerful Multimodal Large Language Models (MLLMs), with their impressive abilities in visual and textual understanding, makes them natural candidates for this task. Consequently, their capabilities in this crucial domain are mostly unquantified. To address this gap, we introduce a new task, Multimodal Interactive Veracity Assessment (MIVA), and present a novel multimodal dataset derived from the social deduction game Werewolf. This dataset provides synchronized video, text, with verifiable ground-truth labels for every statement. We establish a comprehensive benchmark evaluating state-of-the-art MLLMs, revealing a significant performance gap: even powerful models like GPT-4o struggle to distinguish truth from falsehood reliably. Our analysis of failure modes indicates that these models fail to ground language in visual social cues effectively and may be overly conservative in their alignment, highlighting the urgent need for novel approaches to building more perceptive and trustworthy AI systems.