Shadows in the Attention: Contextual Perturbation and Representation Drift in the Dynamics of Hallucination in LLMs

Hallucinations -- plausible yet erroneous outputs -- remain a critical barrier to reliable deployment of large language models (LLMs). We present the first systematic study linking hallucination incidence to internal-state drift induced by incremental context injection. Using TruthfulQA, we construct two 16-round "titration" tracks per question: one appends relevant but partially flawed snippets, the other injects deliberately misleading content. Across six open-source LLMs, we track overt hallucination rates with a tri-perspective detector and covert dynamics via cosine, entropy, JS and Spearman drifts of hidden states and attention maps. Results reveal (1) monotonic growth of hallucination frequency and representation drift that plateaus after 5--7 rounds; (2) relevant context drives deeper semantic assimilation, producing high-confidence "self-consistent" hallucinations, whereas irrelevant context induces topic-drift errors anchored by attention re-routing; and (3) convergence of JS-Drift ($\sim0.69$) and Spearman-Drift ($\sim0$) marks an "attention-locking" threshold beyond which hallucinations solidify and become resistant to correction. Correlation analyses expose a seesaw between assimilation capacity and attention diffusion, clarifying size-dependent error modes. These findings supply empirical foundations for intrinsic hallucination prediction and context-aware mitigation mechanisms.

Robust LLM Fingerprinting via Domain-Specific Watermarks

As open-source language models (OSMs) grow more capable and are widely shared and finetuned, ensuring model provenance, i.e., identifying the origin of a given model instance, has become an increasingly important issue. At the same time, existing backdoor-based model fingerprinting techniques often fall short of achieving key requirements of real-world model ownership detection. In this work, we build on the observation that while current open-source model watermarks fail to achieve reliable content traceability, they can be effectively adapted to address the challenge of model provenance. To this end, we introduce the concept of domain-specific watermarking for model fingerprinting. Rather than watermarking all generated content, we train the model to embed watermarks only within specified subdomains (e.g., particular languages or topics). This targeted approach ensures detection reliability, while improving watermark durability and quality under a range of real-world deployment settings. Our evaluations show that domain-specific watermarking enables model fingerprinting with strong statistical guarantees, controllable false positive rates, high detection power, and preserved generation quality. Moreover, we find that our fingerprints are inherently stealthy and naturally robust to real-world variability across deployment scenarios.

EnSToM: Enhancing Dialogue Systems with Entropy-Scaled Steering Vectors for Topic Maintenance

Small large language models (sLLMs) offer the advantage of being lightweight and efficient, which makes them suitable for resource-constrained environments. However, sLLMs often struggle to maintain topic consistency in task-oriented dialogue systems, which is critical for scenarios such as service chatbots. Specifically, it is important to ensure that the model denies off-topic or malicious inputs and adheres to its intended functionality so as to prevent potential misuse and uphold reliability. Towards this, existing activation engineering approaches have been proposed to manipulate internal activations during inference. While these methods are effective in certain scenarios, our preliminary experiments reveal their limitations in ensuring topic adherence. Therefore, to address this, we propose a novel approach termed Entropy-scaled Steering vectors for Topic Maintenance (EnSToM). EnSToM dynamically adjusts the steering intensity based on input uncertainty, which allows the model to handle off-topic distractors effectively while preserving on-topic accuracy. Our experiments demonstrate that EnSToM achieves significant performance gain with a relatively small data size compared to fine-tuning approaches. By improving topic adherence without compromising efficiency, our approach provides a robust solution for enhancing sLLM-based dialogue systems.

Efficient Online RL Fine Tuning with Offline Pre-trained Policy Only

Improving the performance of pre-trained policies through online reinforcement learning (RL) is a critical yet challenging topic. Existing online RL fine-tuning methods require continued training with offline pretrained Q-functions for stability and performance. However, these offline pretrained Q-functions commonly underestimate state-action pairs beyond the offline dataset due to the conservatism in most offline RL methods, which hinders further exploration when transitioning from the offline to the online setting. Additionally, this requirement limits their applicability in scenarios where only pre-trained policies are available but pre-trained Q-functions are absent, such as in imitation learning (IL) pre-training. To address these challenges, we propose a method for efficient online RL fine-tuning using solely the offline pre-trained policy, eliminating reliance on pre-trained Q-functions. We introduce PORL (Policy-Only Reinforcement Learning Fine-Tuning), which rapidly initializes the Q-function from scratch during the online phase to avoid detrimental pessimism. Our method not only achieves competitive performance with advanced offline-to-online RL algorithms and online RL approaches that leverage data or policies prior, but also pioneers a new path for directly fine-tuning behavior cloning (BC) policies.

VoxRAG: A Step Toward Transcription-Free RAG Systems in Spoken Question Answering

We introduce VoxRAG, a modular speech-to-speech retrieval-augmented generation system that bypasses transcription to retrieve semantically relevant audio segments directly from spoken queries. VoxRAG employs silence-aware segmentation, speaker diarization, CLAP audio embeddings, and FAISS retrieval using L2-normalized cosine similarity. We construct a 50-query test set recorded as spoken input by a native English speaker. Retrieval quality was evaluated using LLM-as-a-judge annotations. For very relevant segments, cosine similarity achieved a Recall@10 of 0.34. For somewhat relevant segments, Recall@10 rose to 0.60 and nDCG@10 to 0.27, highlighting strong topical alignment. Answer quality was judged on a 0--2 scale across relevance, accuracy, completeness, and precision, with mean scores of 0.84, 0.58, 0.56, and 0.46 respectively. While precision and retrieval quality remain key limitations, VoxRAG shows that transcription-free speech-to-speech retrieval is feasible in RAG systems.

MDIT-Bench: Evaluating the Dual-Implicit Toxicity in Large Multimodal Models

The widespread use of Large Multimodal Models (LMMs) has raised concerns about model toxicity. However, current research mainly focuses on explicit toxicity, with less attention to some more implicit toxicity regarding prejudice and discrimination. To address this limitation, we introduce a subtler type of toxicity named dual-implicit toxicity and a novel toxicity benchmark termed MDIT-Bench: Multimodal Dual-Implicit Toxicity Benchmark. Specifically, we first create the MDIT-Dataset with dual-implicit toxicity using the proposed Multi-stage Human-in-loop In-context Generation method. Based on this dataset, we construct the MDIT-Bench, a benchmark for evaluating the sensitivity of models to dual-implicit toxicity, with 317,638 questions covering 12 categories, 23 subcategories, and 780 topics. MDIT-Bench includes three difficulty levels, and we propose a metric to measure the toxicity gap exhibited by the model across them. In the experiment, we conducted MDIT-Bench on 13 prominent LMMs, and the results show that these LMMs cannot handle dual-implicit toxicity effectively. The model's performance drops significantly in hard level, revealing that these LMMs still contain a significant amount of hidden but activatable toxicity. Data are available at https://github.com/nuo1nuo/MDIT-Bench.

Statistical Inference for Online Algorithms

Construction of confidence intervals and hypothesis tests for functionals based on asymptotically normal estimators is a classical topic in statistical inference. The simplest and in many cases optimal inference procedure is the Wald interval or the likelihood ratio test, both of which require an estimator and an estimate of the asymptotic variance of the estimator. Estimators obtained from online/sequential algorithms forces one to consider the computational aspects of the inference problem, i.e., one cannot access all of the data as many times as needed. Several works on this topic explored the online estimation of asymptotic variance. In this article, we propose computationally efficient, rate-optimal, and asymptotically valid confidence regions based on the output of online algorithms {\em without} estimating the asymptotic variance. As a special case, this implies inference from any algorithm that yields an asymptotically normal estimator. We focus our efforts on stochastic gradient descent with Polyak averaging to understand the practical performance of the proposed method.

Learning Interpretable Representations Leads to Semantically Faithful EEG-to-Text Generation

Pretrained generative models have opened new frontiers in brain decoding by enabling the synthesis of realistic texts and images from non-invasive brain recordings. However, the reliability of such outputs remains questionable--whether they truly reflect semantic activation in the brain, or are merely hallucinated by the powerful generative models. In this paper, we focus on EEG-to-text decoding and address its hallucination issue through the lens of posterior collapse. Acknowledging the underlying mismatch in information capacity between EEG and text, we reframe the decoding task as semantic summarization of core meanings rather than previously verbatim reconstruction of stimulus texts. To this end, we propose the Generative Language Inspection Model (GLIM), which emphasizes learning informative and interpretable EEG representations to improve semantic grounding under heterogeneous and small-scale data conditions. Experiments on the public ZuCo dataset demonstrate that GLIM consistently generates fluent, EEG-grounded sentences without teacher forcing. Moreover, it supports more robust evaluation beyond text similarity, through EEG-text retrieval and zero-shot semantic classification across sentiment categories, relation types, and corpus topics. Together, our architecture and evaluation protocols lay the foundation for reliable and scalable benchmarking in generative brain decoding.

Oversmoothing, "Oversquashing", Heterophily, Long-Range, and more: Demystifying Common Beliefs in Graph Machine Learning

After a renaissance phase in which researchers revisited the message-passing paradigm through the lens of deep learning, the graph machine learning community shifted its attention towards a deeper and practical understanding of message-passing's benefits and limitations. In this position paper, we notice how the fast pace of progress around the topics of oversmoothing and oversquashing, the homophily-heterophily dichotomy, and long-range tasks, came with the consolidation of commonly accepted beliefs and assumptions that are not always true nor easy to distinguish from each other. We argue that this has led to ambiguities around the investigated problems, preventing researchers from focusing on and addressing precise research questions while causing a good amount of misunderstandings. Our contribution wants to make such common beliefs explicit and encourage critical thinking around these topics, supported by simple but noteworthy counterexamples. The hope is to clarify the distinction between the different issues and promote separate but intertwined research directions to address them.

ToxicTone: A Mandarin Audio Dataset Annotated for Toxicity and Toxic Utterance Tonality

Despite extensive research on toxic speech detection in text, a critical gap remains in handling spoken Mandarin audio. The lack of annotated datasets that capture the unique prosodic cues and culturally specific expressions in Mandarin leaves spoken toxicity underexplored. To address this, we introduce ToxicTone -- the largest public dataset of its kind -- featuring detailed annotations that distinguish both forms of toxicity (e.g., profanity, bullying) and sources of toxicity (e.g., anger, sarcasm, dismissiveness). Our data, sourced from diverse real-world audio and organized into 13 topical categories, mirrors authentic communication scenarios. We also propose a multimodal detection framework that integrates acoustic, linguistic, and emotional features using state-of-the-art speech and emotion encoders. Extensive experiments show our approach outperforms text-only and baseline models, underscoring the essential role of speech-specific cues in revealing hidden toxic expressions.