LLM 理论进展

Before fixing an issue, it is useful to first reproduce it by generating a bug reproduction test (BRT). However, generating a BRT is itself a challenging task, because issue descriptions tend to be informal, making it difficult to determine whether a candidate BRT indeed fails for the reason in the issue. Prior work has attempted to tackle this problem via inference scaling, using large language models to generate many BRTs and patches, then using execution feedback to select and improve them. Unfortunately, this is expensive and the feedback is unreliable. This paper explores evolutionary programming for BRT generation to sharpen the feedback, while enhancing evolutionary programming to keep costs in check. Our new approach, EvoOtter, controls test execution costs via successive halving. Furthermore, it controls LLM costs via batched crossover for an entire generation in a single LLM call, as well as via rule-based code mutations, with a new fitness score tailored for BRTs. As a result, EvoOtter generates state-of-the-art quality BRTs at the fraction of the cost of prior inference-scaling approaches to this problem. More broadly, this paper points at how to efficiently and effectively combine evolutionary programming with large language models for software engineering.

A central goal of mechanistic interpretability is to understand how neural networks work and what each individual component does. Dominant circuit-finding approaches focus on a specific behavior and reverse-engineer the role of components on the associated sub-distribution. However, past work has shown that components can have different functions that are active on different subsets of the input distribution. In this work we ask whether a single weight can be understood globally across the full training distribution by characterizing when it matters (the inputs on which ablating it changes the model's predictions). We introduce an automated LLM pipeline that writes a short, human-readable description of when a weight matters and verifies it on held-out text, crediting a weight only if its description generalizes. Across two sparse and two dense transformers, the fraction of weights that are interpretable (in this sense) is higher in sparse transformers than in dense ones, a gap that widens once unreliable descriptions are discarded. Our results show that a meaningful fraction of a sparse transformer model's weights can be interpreted: 12 to 31% of weights have a single short description that identifies what the weight is used for.

Large language models remain limited as continual learning systems, motivating renewed interest in Sparse Distributed Memory (SDM) as an explicit online episodic memory. CALM (Nechesov and Ruponen, 2025) identifies its threshold-binary encoder as an open design question. This paper evaluates rank-order N-of-M encoding (Furber et al., 2007) as an alternative. We make three contributions. First, a faithful reimplementation validates the published architecture by confirming exact equivalence between WheelSDM and RankOrderSDM (cosine similarity 1.0000 across 10 seeds) and reproducing the documented divergence of RDLIF neurons under interference. Second, multi-seed capacity experiments show RankOrderSDM outperforming StandardSDM by 13.4 percentage points at saturation in the scaled configuration and by 0.8 percentage points at the published architecture scale. Third, BER robustness experiments disentangle representation and learning effects, showing that the large robustness gain arises primarily from the interaction of rank-order encoding with MAX-Hebbian learning, while the encoder alone provides only a small advantage under matched learning conditions. Experiments on GloVe-100 embeddings confirm this small but consistent encoding benefit on real structured data, whereas sentence embeddings exhibit a ceiling effect at low memory load. A secondary analysis shows that idealized rank-order encoding requires half the component-level encoding energy of SpikingMamba's SI-LIF neurons at four-bit precision, although decoder costs dominate overall system energy. These results identify which components of the original rank-order SDM architecture provide measurable benefits for contemporary memory-augmented AI systems, offering practical guidance for architectures such as CALM.

When a person shares something with a language model, the model often answers the surface of the message rather than what the sender was doing by sending it: share a finished project and it critiques the code; share a raw late-night line and it runs a wellness check. We treat the sender's communicative intent, the Gricean what-was-meant, as a first-class interpretability object, and show the failure is one of readout on top of a robust representation. A linear probe decodes the sender's intent, whether they want a thing recognized or evaluated, from a model's default-pass hidden states, cleanly and surface-independently, across six models and four families and in the base checkpoints. The representation generalizes further, to intent that is only pragmatically inferred, and to a second, lexically clean intent (support versus help). The behavioral half of the story, and every causal test, is established on the recognize/evaluate contrast, where what varies is whether the default output acts on the intent. The readout lags the representation in depth within a model (the intent is decodable several layers before it drives the output); across models, which ones act on it by default is model-specific, an observed stratification (three of six show the failure) that we do not read as a scaling law. Where the gap is open, a direction closely tied to the representation, the discriminative direction at a searched-for layer, is a causal handle: steering it recovers the intended behavior, as well as an explicit instruction does and with no prompt at all. This direction is near-orthogonal to the feedback-offering axis, so it routes a represented intent rather than a generic feedback knob, though at the recovery dose the routed intent can override an explicit request. We support each link with controls against obvious deflations and report the nulls as plainly as the confirmations.

Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate chunk selection. We propose Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query performs attention independently with each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training. Experimental results show that HiLS-Attention achieves performance comparable to, and in some cases better than, full attention at in-domain context lengths. Meanwhile, HiLS-Attention extrapolates more than $64\times$ the training context length with 90% retrieval accuracy, far beyond full attention. Moreover, existing full-attention models can be converted to HiLS-Attention with lightweight continued pretraining, preserving in-domain performance while acquiring ultra-long-context extrapolation. Together with its sparse KV access and computation, HiLS-Attention breaks the usual efficiency-performance trade-off, enabling long-context LLMs that are both more efficient and more effective on general long-context tasks than their full-attention counterparts.

While Large Language Models (LLMs) can provide personalized support in learning, several studies have raised concerns regarding their use in education. Importantly, learning depends on how students engage with LLMs. This study examined how two types of LLM-based tutors shape students' prompting practices, learning, and subsequent LLM-use: a Socratic-Guidance (SG) tutor, which structures interaction through dialogic questioning, and a Prompt-Refinement (PR) tutor that guides the formulation of effective prompts. We conducted a two-phase study in a graduate-level mobile robotics course: 66 students used either the SG or PR tutor during a 6-week intervention, followed by 52 students using an unconstrained LLM during a 3-week course project. Results show that while the SG- and PR tutors led to similar task performance and prompting patterns during guided use, they differ in learning outcomes and later LLM-use. SG-students, relative to PR-student, achieved higher learning gains in later sessions, and were more likely to adopt understanding-driven prompting strategies, which are predictive of higher understanding, when using an unconstrained LLM. Although learners perceived the SG tutor as less efficient, the findings suggest that Socratic guidance supports the development of students' capacity to learn with LLMs over time, highlighting its importance for LLM tutor design.

Thinking has emerged as a critical capability for Large Language Models (LLMs) tackling complex tasks. However, its reactive nature, where reasoning is passively triggered only upon receiving a user response, inevitably introduces latency that compromises conversational fluidity. This stands in sharp contrast to human dialogue, where speakers proactively anticipate and plan future content during natural pauses to ensure seamless interaction. To bridge this gap, we propose Proactive Thinking, a framework that empowers models to pre-compute potential response elements during conversational downtime instead of waiting idly for the next input. We then introduce a training-free baseline that can think ahead by anticipating future states, balancing efficiency and quality through speculative continual thinking. To evaluate this approach in practice, we adapt three benchmarks of varying complexity into time-aware environments that simulate real-time conversational flow. We demonstrate that proactive thinking effectively improves interaction efficiency without compromising performance. Ultimately, this work advocates for a fundamental shift toward more intelligent, anticipatory, and real-time conversational AI.

Social media posts are a rich and valuable source of data for analyzing mental health states and users' well-being using automated analysis tools. In this work, we demonstrate how we used a range of Natural Language Processing (NLP) methods, including Long Short-Term Memory (LSTM), BERT-based models, and Large Language Models (LLMs), for self-state and well-being analysis and summarization during the CLPsych Shared Task 2026. Our approach achieved one of the top Consistency and Contradiction scores for the summarization task and also middle-level results for the other tasks. By testing and developing such mental health-state estimation systems, we contributed to improving mental health support systems. We make our code available https://github.com/psytechlab/CLPsych2026/.

This paper presents an approach to the SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis. We investigate methods for moving beyond traditional categorical sentiment (e.g., positive or negative) to predict fine-grained, real-valued scores for sentiment "valence" (positivity) and "arousal" (intensity). We participate in two subtasks: predicting these scores for given aspects (Subtask 1) and extracting full sets of sentiment details, including aspects, categories, and opinions alongside their scores (Subtask 3). Our approach for the regression task involves a weighted ensemble of transformer-based encoder models. For the Russian language, we further enhance the input by using a large language model (LLM) to generate synthetic sentiment descriptions. For the extraction task, we fine-tune a decoder LLM to perform structured prediction, allowing the system to identify sentiment elements and estimate their numerical scores simultaneously.

This study examines how prompt language and translation theory-driven prompt design influence the quality of Spanish-Chinese journalistic translations generated by GPT-5.2. A parallel corpus of four editorials from El Pais was translated under 48 experimental conditions (4 prompt types, 3 prompt languages, and 4 articles). Translation quality was assessed using BLEU and BERTScore-F1 for automated evaluation, alongside human evaluation based on the Multidimensional Quality Metrics (MQM) framework. Automated metrics identified the baseline prompt (BASE) as the best-performing condition, whereas human evaluation ranked the brief-oriented prompt (BRIEF) highest (MQM: 8.66 vs. 7.84), a reversal likely attributable to the single-reference constraint inherent in automated measures. Sub-error type analysis revealed that translation theory-driven prompts selectively reduced Awkward style errors, while Unidiomatic style errors persisted across conditions. Prompt language had a negligible impact under both evaluation paradigms. These results indicate that translation theory-driven prompts can yield measurable quality gains under expert evaluation of journalistic translations, although their pedagogical implications for language learners remain suggestive and require validation through user-based studies.

Malicious Python packages have become a major threat to software supply chain ecosystems due to the widespread adoption of open-source repositories such as PyPI. Existing learning-based detection methods struggle to capture the hierarchical organization and heterogeneous interactions among different program entities. Although Large Language Models (LLMs) have demonstrated strong capabilities in code understanding and semantic reasoning, they are rarely integrated with structural program representations for fine-grained malicious behavior analysis. In this paper, we propose an LLM-enhanced hierarchical heterogeneous graph representation learning framework for malicious Python package detection. The framework constructs a hierarchical heterogeneous code graph that explicitly models heterogeneous code entities and different types of structural dependencies. LLMs are further leveraged to infer function-level semantic roles, introducing an additional layer of semantic heterogeneity. Based on this graph, we develop a hierarchical heterogeneous graph neural network that performs type-aware message passing over different node and edge categories, effectively modeling malicious behavior propagation for accurate package-level classification. The framework also incorporates a function-level attribution mechanism which, combined with LLM reasoning, automatically identifies suspicious functions and localizes fine-grained malicious behaviors without human expert intervention. Extensive experiments on real-world datasets show that our framework consistently outperforms traditional machine learning methods, graph-based detectors, and state-of-the-art LLMs across packages with varying sizes and dependency complexities, while providing accurate, robust, and interpretable malicious behavior localization.

We present an automated pipeline that decomposes Italian tax-court judgments into individual legal issues and extracts, for each issue, a structured XML representation grounded in the IRAC framework and the legal syllogism. The pipeline targets a corpus of approximately $330{,}000$ first- and second-instance decisions of the Italian tax courts and is built around a capable yet cost-efficient general-purpose model (DeepSeek V3), a choice driven by the need to process several hundred thousand documents at a sustainable cost. To address the well-documented unreliability of large language models on legal citations, we couple the extraction step with an automatic hallucination-detection filter that compares the references produced by the model with those identified in the judgment text by a dedicated parser (Linkoln), normalised to standard identifiers (URN-NIR, ECLI, CELEX). We validate the pipeline on $50$ judgments annotated by two PhDs in tax law, computing inter-annotator agreement and LLM-vs-expert agreement on both issue extraction and legal citations, together with a stand-alone evaluation of the hallucination filter. To the best of our knowledge, this is the first issue-level, expert-validated structured extraction pipeline with hallucination control for Italian tax-court decisions, and it provides a concrete starting point for downstream applications such as issue-level retrieval, citation-network analysis, and the construction of large-scale datasets of legal reasoning.

Inference-time alignment methods, such as Best-of-$N$, offer a flexible alternative to training-based alignment by using reward models to select high-quality responses generated by a reference LLM. However, the efficacy of these methods is inherently limited by the response quality: if the reference LLM assigns negligible probability to high-reward responses, no selection strategy will succeed in finding aligned outputs. In this work, we propose Best-of-Better-$N$ (BoBN), an in context learning-based generation framework to address this challenge. Our method utilizes retrieval from high-reward examples relevant to the input query and task. Crucially, we introduce a restyling step where retrieved responses are rewritten by the reference LLM to align with the target task's format and style. These restyled examples are used in-context to shift the sampling distribution toward the high-reward region. We analytically characterize how in-context learning shifts the output distribution of pretrained transformers toward the high-reward region, resulting in provable benefits on the target task. We then evaluate BoBN on safety alignment and mathematical reasoning benchmarks across several reference LLMs. BoBN's higher-quality responses enable better performance to be achieved when the number of responses $N$ is fixed, and smaller $N$ required to achieve a target performance.

Since 2020, the Linux Foundation and the multi-organization Inclusive Naming Initiative (INI) have encouraged open-source projects to replace non-inclusive terms such as master/slave and whitelist/blacklist. Although these recommendations have been widely adopted, there is limited empirical evidence on their long-term adoption across Linux Foundation (LF) projects or their implications for AI-assisted software development. In this paper, we present NISCAN, a multilingual static-analysis framework that detects non-inclusive terminology across source code and related software artifacts using the INI vocabulary. Using NISCAN, we conduct the first ecosystem-scale study of inclusive naming across 461 Linux Foundation repositories. Our analysis shows that non-inclusive terminology has declined by approximately 47% since 2020, yet adoption remains incomplete: 62.7% of repositories still contain at least one Tier-1 non-inclusive identifier, while most remaining terminology resides outside source code in documentation, comments, configuration files, and other software artifacts. We further show that repository size, programming language, project functionality, and ecosystem are stronger predictors of term inclusiveness in LF repositories rather than foundation governance. To examine the implications for AI-assisted software development, we conduct a case study evaluating whether large language models (LLMs) can reconstruct legacy non-inclusive identifiers from surrounding program context. The results show that historical naming decisions remain embedded in model predictions even after identifiers have been renamed. Overall, our study findings provide the first ecosystem-scale assessment of inclusive naming adoption within the Linux Foundation and highlight the importance of addressing terminology residue to support responsible naming and ethically sourced code generation.

We introduce Gemma 4, a new generation of open-weight, natively multimodal language models in the Gemma model family. Designed to advance compute efficiency and reasoning, the Gemma 4 model suite features dense and Mixture-of-Experts architectures, ranging from 2.3B to 31B parameters. Alongside improved vision and audio encoders for all model sizes, we propose a unified, encoder-free architecture for our 12B model, which ingests raw audio and image patches. Furthermore, we integrate a thinking mode, enabling Gemma models to generate reasoning traces prior to responding. We improve inference speed, memory, and compute efficiency, as well as long-context abilities through critical design choices. Gemma 4 establishes a leap in performance across STEM, multimodal, and long-context benchmarks, and rivals larger, frontier open models in human-rated tasks.

N-Version Programming (NVP) improves software reliability by executing multiple independent implementations and combining outputs, but its adoption is limited by high cost and the assumption of failure independence, which empirical studies have challenged. Recent advances in Large Language Models (LLMs) reduce the cost of generating multiple implementations, shifting focus to whether their failures are independent. We propose the first systematic methodology to assess failure independence in LLM-generated code and apply it to 224 problems across twelve models, five languages, and three prompting strategies. We analyze both structural and behavioral diversity (i.e., whether implementations fail on the same test cases), complemented by N-version reliability analysis under majority voting and manual inspection of the generated code. Structural diversity analysis shows that implementations from the same model are highly similar, while different models produce more distinct solutions. The same trend appears in behavioral diversity, with implementations from different models showing higher diversity yet still failing on the same tests far more often than expected under independence. N-version reliability analysis reinforces this: three- and five-version ensembles realize only 0.43 and 0.44 of the reliability gain achievable under independence, dropping below 0.3 when ensembles are built from the same model. Manual fault analysis shows that even different failure patterns often share root causes. Overall, these results suggest LLM-generated solutions do not satisfy NVP's failure independence assumption, though heterogeneous models help partially. They also validate our methodology as a tool for systematically evaluating failure independence as models evolve.

High-throughput long-context generation is one of the central challenges for large language models. Generation is typically memory-bandwidth-bound rather than compute-bound: each decoding step must stream the accumulated key/value (KV) cache from memory, so bandwidth demand grows with context length while only one token is emitted. Two parallel approaches have therefore emerged: reducing memory access with efficient attention variants and linear-time mixers such as Mamba, or increasing parallel computation by generating blocks of tokens at once. However, technical challenges arise when combining these two ideas. Earlier hybrid diffusion models such as DiffuMamba use bidirectional Mamba mixing, including a reverse-direction scan relative to causal generation. This reverse scan needs to scan the entire sequence, so its states are not prefix-only and cannot be precisely reused as a cache even when diffusion is performed block by block. We propose a BDLM Mamba--attention hybrid that addresses this challenge by restricting the reverse Mamba scan to the active denoising block, which enables exact caching across blocks. In an 87M-parameter DCLM sweep, BDLM Mamba-H achieves the best C4-en validation perplexity compared to BDLM attention and full-sequence baselines. At 350M parameters, it remains competitive with BDLM attention. For long-context inference, BDLM Mamba-H reaches 19.7x the throughput of full-sequence DiffuMamba-H at 65K tokens and 3.7x the throughput of BDLM attention at 262K, showing that Mamba hybrids are a potential long-context diffusion architecture.

Induction heads are attention circuits believed to underlie in-context learning in transformers, yet a precise characterization of the estimators they implement remains elusive. We study transformers trained on order-$k$ Markov chains and identify two complementary smoothing mechanisms. First, at finite attention-weight scale, the circuit implements a soft context-matching estimator: it aggregates contributions from exact and partial context matches, weighted exponentially by their overlap, and induces a data-dependent interpolation across context orders analogous to Jelinek-Mercer smoothing. Second, a beginning-of-sequence (BOS) token induces additive pseudo-counts, recovering Dirichlet-style smoothing. We construct a disentangled transformer implementing both mechanisms and show that trained transformers recover the predicted attention patterns. Across settings where pseudo-count smoothing is optimal or lower-order contexts provide structured evidence, trained transformers match or outperform classical count-based baselines. Our results bridge mechanistic interpretability of induction heads with classical statistical smoothing, revealing that transformers learn to regularize in-context estimation rather than simply count.

Diffusion Language Models (DLMs) have recently emerged as a promising alternative to autoregressive models. Unlike standard diffusion-based approaches, DLMs are not explicitly conditioned on a timestep, raising a natural question: do these models internally represent denoising progress, and how is such information used downstream? In this work, we show that DLMs do in fact encode a latent representation related to the diffusion timestep within their residual streams. We find that this signal can be reliably extracted using probes across layers, indicating that denoising progress is decodable from internal activations. We further demonstrate that steering the model along a low-dimensional subspace associated with the inferred timestep allows us to systematically modulate its notion of denoising progress, leading to predictable changes in model confidence and entropy. Finally, we analyse the geometry of the identified representation, showing that it exhibits structured and interpretable properties in activation space, and shedding light on how such a signal is processed by these models.

Chemical language models (CLMs) are trained with linearized representations such as SMILES, yet it remains unclear which chemically meaningful substructures they encode. To foster a better understanding of CLMs, we conduct a systematic study and probe for 78 molecular substructures across eight pre-trained and six randomly initialized models. We furthermore study how fine-tuning on chemical downstream tasks affects the learned representations of molecular substructures. Our results show that pre-training generally improves molecular structure awareness of CLMs, particularly in the upper layers. Moreover, randomly initialized models already encode ring structures well in the first layer. Our analysis on two chemical downstream tasks further reveals that, interestingly, fine-tuning affects task-relevant molecular substructures more than others, indicating that the changes in the representations follow chemical theory.

While Vision Transformers have achieved remarkable success across computer vision and language applications, the geometric evolution of their internal representations throughout training remains insufficiently understood. Existing analyses primarily focus on attention mechanisms and downstream performance, leaving the evolution of representation geometry largely unexplored. In this work, we present Transformer Geometry Observatory-II (TGO-II), a representation geometry analysis framework designed to investigate how Transformer representations evolve during supervised training. TGO-II analyzes Vision Transformer (ViT-Small/16) representations using Centered Kernel Alignment (CKA), Singular Vector Canonical Correlation Analysis (SVCCA), Two-Nearest Neighbor Intrinsic Dimensionality (TwoNN-ID), and token covariance analysis. Our experiments reveal three key observations. First, both CKA and SVCCA progressively decrease throughout training, indicating increasing representational specialization across Transformer layers. Second, intrinsic dimensionality consistently increases before stabilizing, suggesting progressive expansion of the representation manifold into a larger set of locally accessible degrees of freedom. Third, token covariance and coupling analyses demonstrate that strong token interaction structure persists throughout training, challenging the hypothesis that increasing representational complexity arises primarily from progressive token independence. These findings suggest that representation complexity and layer specialization emerge simultaneously during training. Manifold expansion appears to occur without token decoupling. Together, these observations motivate a new hypothesis in which Vision Transformers increase representational complexity through progressively richer transformations while preserving strong token interaction structure during learning.

Natural Language Processing (NLP) has traditionally been published in its core disciplinary venues like ACL. However, advances in Large Language Models (LLMs) has led to a blurring of the disciplinary lines between NLP and general Machine Learning (ML), with authors regularly publishing in venues from both fields. Here, we ask whether the disciplinary center of gravity is shifting. Using NLP research published from 2010 to 2026 and studies of both established and new authors, we find that a migration is taking place. First, comparing the pre- and post-LLM eras, established authors lost 19.2pp of share at flagship *ACL main-conference tracks while gaining 14.8pp in the newer Findings tracks, and general ML venues rose 8.6pp, even when adjusting for parallel growth in the fields. Second, among newer authors who debut with at least three first-author NLP-topic papers, the share whose work appears mostly at *ACL venues fell from 84% (2019) to 74% (2024), while the share appearing mostly at general ML venues rose from 5% to 21%. Using causal inference techniques, we estimate that these general ML venues confer a significant citation premium, which influences venue selection. Together, these results point to a significant shift in where NLP research is published.

Language models are increasingly used to quantify cultural phenomena, but what makes such measurement distinctively cultural? This paper argues that NLP work on culture is a material-discursive practice: the apparatus — model, data, annotation, evaluation — participates in constituting the cultural reality it measures, rather than passively recording it. Drawing on Karen Barad's concept of the agential cut — the contingent boundary between phenomenon and instrument — I show that the apparatus's substantive design choices draw such boundaries, and that the boundary is entangled from the start because language models have already internalized much of the cultural material they measure. I illustrate this through three case studies on television and film dialogue (measuring structure, interaction, and deviation) and three examinations of the apparatus itself (erasure of cultural markers, attunement to historical material, and agency in an agentic workflow). This big picture analysis proposes a research program that is theory-driven, empirically rigorous, and culturally contingent, treating each agential cut as a conscious commitment, at once methodological and ethical.

Safety training for large language models (LLMs) is conducted predominantly in English, leaving uncertain how well safety mechanisms generalize to low-resource languages and mixed-language code-switching. We show that this creates an epistemic gap in which models confidently generate harmful responses for inputs that fall outside the distribution of their safety training. To study this phenomenon, we introduce STEER (Safety Targeted Embedding Exploit via Refinement), a gradient-guided attack that identifies words contributing most strongly to the model's refusal behavior and iteratively translates them into low-resource languages to suppress refusal while preserving harmful intent. Across six open-source 8B-parameter models, STEER achieves attack success rates of up to 93.0% on JailbreakBench and 96.7% on AdvBench, outperforming random code-switching and Greedy Coordinate Gradient (GCG). The resulting prompts also transfer to GPT-4o-mini, achieving a 35.5% attack success rate without requiring access to the target model, suggesting that the underlying weakness is not specific to a single architecture. These findings demonstrate that safety mechanisms aligned primarily on English cannot be assumed to generalize across multilingual inputs. We argue that improving multilingual safety requires broader coverage during alignment and mechanisms that explicitly detect and abstain on out-of-distribution inputs.

Online multimodal knowledge editing requires injecting a continual stream of visual-textual corrections into multimodal large language models (MLLMs) with bounded overhead and minimal disruption to unrelated behaviors. Existing editors mainly emphasize edit reliability and long-horizon stability, but rarely control the semantic boundary of each edit. Our pilot analyses of post-edit behaviors and internal neuronal activities reveal a scope gap behind reliable edits: instance-level success neither guarantees transfer to valid cross-modal variants nor prevents leakage to unrelated inputs, while edit-related cross-modal responses concentrate in deeper semantic layers. Therefore, we formulate Edit-Scoped Generalization, reframing online MLLM editing from merely correcting an instance to controlling the propagation boundary of each edit. To this end, we propose ScopeEdit, a scope-aware online editor that decomposes each update into a modality-local absorption branch and an evidence-gated shared generalization branch. The local branch supports stable edit absorption, whereas the shared branch enables cross-modal propagation only when visual and textual evidence are sufficiently aligned. Both branches perform scope-separated write geometries in orthogonal low-rank spaces and maintain branch-wise preconditioners via Sherman--Morrison recursions, yielding constant per-edit overhead. Extensive experiments across diverse benchmarks, long-horizon edit streams, MLLM backbones, real-world VLKEB scenarios, and complex vision-language architectures show that ScopeEdit consistently improves the trade-off between in-scope cross-modal transfer and out-of-scope locality, while preserving edit reliability, stability and online efficiency. Our code is available at https://github.com/lab-klc/ScopeEdit.

In this paper, we introduce BamiBERT, a new BERT-based pre-trained language model for Vietnamese that addresses key limitations of PhoBERT — the current de facto Vietnamese text encoder. Trained from scratch on a 129GB corpus of general-domain Vietnamese text for 20 epochs, BamiBERT supports an extended context length of up to 2048 tokens and operates directly on raw input, eliminating the need for external word segmentation. Across 8 Vietnamese benchmarks, it achieves the best score on 11 of 15 metrics and the second-best on 3 others, setting a new state of the art among "base"-sized Vietnamese encoders and demonstrating strong cross-domain generalization. We release BamiBERT at: https://huggingface.co/Qualcomm-AI-Research/BamiBERT

Large language model personas are increasingly used to approximate diverse users during early-stage visualization design, but it remains unclear whether persona-conditioned outputs reflect stable personality effects or artifacts of model choice and task framing. We examine this question across two visualization-relevant tasks: color assignment for abstract and concrete concepts, and chart-idiom preference ratings across task contexts. Using 43 Big Five profiles across GPT-4o-mini, GPT-4.1-mini, and GPT-5-mini, we find that personality-color coupling is highly model-configuration dependent: absent in GPT-4o-mini for all six concepts, consistent in GPT-4.1-mini across all six, and partial in GPT-5-mini for two of six. Concept type further shapes the signal: for abstract concepts, personality explains more hue variance than model identity, while concrete concepts show smaller and comparable effects. In chart choice, trait-aligned cluster aggregation produces stable top-idiom rankings across all nine cluster-context combinations, but a no-persona baseline recovers the same top choice in 8 of 9 model-context cells, indicating that task context drives rank-1 selection more than personality. These findings position LLM personas as exploratory probes for visualization design, not substitutes for human participants, and motivate multi-model testing, concept-type disaggregation, and no-persona baselines in future studies.

Large Language Models (LLMs) have recently shown promise in molecular discovery, yet a gap remains between their probabilistic nature over discrete sequential tokens and the rigid topological constraints of chemical space. This raises the question of whether molecular LLMs can generalize beyond the local neighborhoods induced by their sequence-based representations. To systematically investigate this question, we introduce a Molecular Perturbation framework that generates syntax-valid structural variants of training molecules under controlled Graph Edit Distance (GED) to probe the manifold regularity of molecular LLMs. Our analysis shows that even a single edit can cause substantial performance drops on common molecular tasks, revealing a narrow local trust region and fragile sensitivity to structural changes. Since similar molecules tend to exhibit similar properties, In-Context Tuning (ICT), which anchors predictions on structurally similar molecules, offers a natural way to mitigate such fragility. Our experiments also examine whether ICT confers robustness under controlled structural perturbations, and the results suggest that it can partially expand the local trust region and offer a promising direction for stabilizing molecular LLMs against structural variation.

As Large Language Models (LLMs) and agentic systems become integrated into real-world applications, ensuring their safety and security is critical. Guardrail systems that detect and block malicious instructions sent to and from an LLM are an essential component of AI security. However, researchers conducting black-box adversarial emulation against production AI systems often struggle to determine whether a guardrail block or an LLM rejection has occurred. This distinction is important because the techniques used to bypass guardrails can differ substantially from those used to bypass LLM safety alignment, and has a material impact on attack technique selection and optimization. We propose the first black-box guardrail reconnaissance methodology, which detects the presence of a guardrail within a target AI system through behavioral monitoring of HTTP, lexical, and timing signals, assuming only black-box access and zero prior knowledge of the guardrail or AI system. Experiments demonstrate that our approach detects guardrail presence with 100% accuracy, with statistically significant behavioral separation between benign and malicious interactions (q < 0.001). Our approach further identifies the content categories a guardrail is designed to block, and distinguishes guardrail blocks from LLM rejection on unseen prompts with an average F1 score of 98%.

Large language models (LLMs) are increasingly consulted on contested scientific questions, raising the concern that they will sycophantically retreat from established consensus when a user signals doubt — drifting toward a false balance that treats settled science as one view among several. We test this across three open instruction-tuned models (Llama-3.1-8B, Qwen2.5-7B, Mistral-7B), three consensus-science domains (climate, vaccines, evolution), and single- and multi-turn settings, combining behavioral measurement with linear probing and activation patching. We do not observe sycophantic retreat. Instead, models show three distinct policies under the same skeptical pressure: reactive assertion, where consensus assertion increases rather than decreases (Llama); surface hedging, where tone softens while the position holds (Qwen); and non-response (Mistral). Pairwise judgments confirm the reactive shift is stance, not style (63.6%, p=.007), and a decomposition identifies increased consensus assertion, not false balance, as its driver (beta=+0.042 per dose, p<1e-77). Linear probes localize the divergence to middle layers — perfect separation in Llama and Qwen versus 72% in Mistral, with non-overlapping confidence intervals — indicating the non-responsive model does not linearly represent the skepticism signal at all. Crucially, this robustness does not transfer: it attenuates across domains and, in the safety-critical vaccine domain, can reverse, with myth-rebuttal weakening under skeptical pressure. We synthesize these into a four-way taxonomy separating active from accidental robustness, and argue that behavioral evaluation alone cannot distinguish a model that resists skepticism because it understands the signal from one that only appears to resist because it fails to perceive it.