Modern LLMs face inference efficiency challenges due to their scale. To address this, many compression methods have been proposed, such as pruning and quantization. However, the effect of compression on a model's interpretability remains elusive. While several model interpretation approaches exist, such as circuit discovery, Sparse Autoencoders (SAEs) have proven particularly effective in decomposing a model's activation space into its feature basis. In this work, we explore the differences in SAEs for the original and compressed models. We find that SAEs trained on the original model can interpret the compressed model albeit with slight performance degradation compared to the trained SAE on the compressed model. Furthermore, simply pruning the original SAE itself achieves performance comparable to training a new SAE on the pruned model. This finding enables us to mitigate the extensive training costs of SAEs.

Recently, Zaremba et al. demonstrated that increasing inference-time computation improves robustness in large proprietary reasoning LLMs. In this paper, we first show that smaller-scale, open-source models (e.g., DeepSeek R1, Qwen3, Phi-reasoning) can also benefit from inference-time scaling using a simple budget forcing strategy. More importantly, we reveal and critically examine an implicit assumption in prior work: intermediate reasoning steps are hidden from adversaries. By relaxing this assumption, we identify an important security risk, intuitively motivated and empirically verified as an inverse scaling law: if intermediate reasoning steps become explicitly accessible, increased inference-time computation consistently reduces model robustness. Finally, we discuss practical scenarios where models with hidden reasoning chains are still vulnerable to attacks, such as models with tool-integrated reasoning and advanced reasoning extraction attacks. Our findings collectively demonstrate that the robustness benefits of inference-time scaling depend heavily on the adversarial setting and deployment context. We urge practitioners to carefully weigh these subtle trade-offs before applying inference-time scaling in security-sensitive, real-world applications.

Alignment of large language models (LLMs) typically involves training a reward model on preference data, followed by policy optimization with respect to the reward model. However, optimizing policies with respect to a single reward model estimate can render it vulnerable to inaccuracies in the reward model. We empirically study the variability of reward model training on open-source benchmarks. We observe that independently trained reward models on the same preference dataset can exhibit substantial disagreement, highlighting the instability of current alignment strategies. Employing a theoretical model, we demonstrate that variability in reward model estimation can cause overfitting, leading to the risk of performance degradation. To mitigate this risk, we propose a variance-aware policy optimization framework for preference-based alignment. The key ingredient of the framework is a new policy regularizer that incorporates reward model variance estimates. We show that variance-aware policy optimization provably reduces the risk of outputting a worse policy than the default. Experiments across diverse LLM and reward model configurations confirm that our approach yields more stable and robust alignment than the standard (variance-unaware) pipeline.

Empowered by large language models (LLMs), intelligent agents have become a popular paradigm for interacting with open environments to facilitate AI deployment. However, hallucinations generated by LLMs--where outputs are inconsistent with facts--pose a significant challenge, undermining the credibility of intelligent agents. Only if hallucinations can be mitigated, the intelligent agents can be used in real-world without any catastrophic risk. Therefore, effective detection and mitigation of hallucinations are crucial to ensure the dependability of agents. Unfortunately, the related approaches either depend on white-box access to LLMs or fail to accurately identify hallucinations. To address the challenge posed by hallucinations of intelligent agents, we present HalMit, a novel black-box watchdog framework that models the generalization bound of LLM-empowered agents and thus detect hallucinations without requiring internal knowledge of the LLM's architecture. Specifically, a probabilistic fractal sampling technique is proposed to generate a sufficient number of queries to trigger the incredible responses in parallel, efficiently identifying the generalization bound of the target agent. Experimental evaluations demonstrate that HalMit significantly outperforms existing approaches in hallucination monitoring.

Small language models have gained significant popularity due to their efficiency and growing capabilities. However, incorporating additional modalities, such as vision, can exacerbate the challenge of limited context windows by introducing noise. Recent studies have highlighted that Transformer attention mechanisms often disproportionately focus on irrelevant contexts. In this work, we extend the Differential Attention mechanism, originally designed for text-only models, to the text-vision model PaliGemma. Our aim is to evaluate its ability to mitigate noisy information retrieval and reduce hallucinations. To this end, we fine-tuned the PaliGemma 3B model using LoRA, incorporating Differential Attention, and experimented with various parameter settings and configurations. We demonstrate that Differential Attention can be adapted and integrated into the fine-tuning of existing models to enhance noisy information retrieval and question-answering capabilities.

Large language models (LLMs) now write code in settings where misreading a single word can break safety or cost money, yet we still expect them to overlook stray typos. To probe where useful robustness ends and harmful insensitivity begins, we compile 50 LeetCode problems and craft three minimal prompt perturbations that should vary in importance: (i) progressive underspecification deleting 10 % of words per step; (ii) lexical flip swapping a pivotal quantifier ("max" to "min"); and (iii) jargon inflation replacing a common noun with an obscure technical synonym. Six frontier models, including three "reasoning-tuned" versions, solve each mutated prompt, and their Python outputs are checked against the original test suites to reveal whether they reused the baseline solution or adapted. Among 11 853 generations we observe a sharp double asymmetry. Models remain correct in 85 % of cases even after 90 % of the prompt is missing, showing over-robustness to underspecification, yet only 54 % react to a single quantifier flip that reverses the task, with reasoning-tuned variants even less sensitive than their bases. Jargon edits lie in between, passing through 56 %. Current LLMs thus blur the line between harmless noise and meaning - changing edits, often treating both as ignorable. Masking salient anchors such as function names can force re - evaluation. We advocate evaluation and training protocols that reward differential sensitivity: stay steady under benign noise but adapt - or refuse - when semantics truly change.

Rare diseases affect 1 in 10 Americans, yet standard ICD coding systems fail to capture these conditions in electronic health records (EHR), leaving crucial information buried in clinical notes. Current approaches struggle with medical abbreviations, miss implicit disease mentions, raise privacy concerns with cloud processing, and lack clinical reasoning abilities. We present Rare Disease Mining Agents (RDMA), a framework that mirrors how medical experts identify rare disease patterns in EHR. RDMA connects scattered clinical observations that together suggest specific rare conditions. By handling clinical abbreviations, recognizing implicit disease patterns, and applying contextual reasoning locally on standard hardware, RDMA reduces privacy risks while improving F1 performance by upwards of 30\% and decreasing inferences costs 10-fold. This approach helps clinicians avoid the privacy risk of using cloud services while accessing key rare disease information from EHR systems, supporting earlier diagnosis for rare disease patients. Available at https://github.com/jhnwu3/RDMA.

We present the DEREK (Deep Extraction & Reasoning Engine for Knowledge) Module, a secure and scalable Retrieval-Augmented Generation pipeline designed specifically for enterprise document question answering. Designed and implemented by eSapiens, the system ingests heterogeneous content (PDF, Office, web), splits it into 1,000-token overlapping chunks, and indexes them in a hybrid HNSW+BM25 store. User queries are refined by GPT-4o, retrieved via combined vector+BM25 search, reranked with Cohere, and answered by an LLM using CO-STAR prompt engineering. A LangGraph verifier enforces citation overlap, regenerating answers until every claim is grounded. On four LegalBench subsets, 1000-token chunks improve Recall@50 by approximately 1 pp and hybrid+rerank boosts Precision@10 by approximately 7 pp; the verifier raises TRACe Utilization above 0.50 and limits unsupported statements to less than 3%. All components run in containers, enforce end-to-end TLS 1.3 and AES-256. These results demonstrate that the DEREK module delivers accurate, traceable, and production-ready document QA with minimal operational overhead. The module is designed to meet enterprise demands for secure, auditable, and context-faithful retrieval, providing a reliable baseline for high-stakes domains such as legal and finance.

Video Object Segmentation (VOS) is a core task in computer vision, requiring models to track and segment target objects across video frames. Despite notable advances with recent efforts, current techniques still lag behind human capabilities in handling drastic visual variations, occlusions, and complex scene changes. This limitation arises from their reliance on appearance matching, neglecting the human-like conceptual understanding of objects that enables robust identification across temporal dynamics. Motivated by this gap, we propose Segment Concept (SeC), a concept-driven segmentation framework that shifts from conventional feature matching to the progressive construction and utilization of high-level, object-centric representations. SeC employs Large Vision-Language Models (LVLMs) to integrate visual cues across diverse frames, constructing robust conceptual priors. During inference, SeC forms a comprehensive semantic representation of the target based on processed frames, realizing robust segmentation of follow-up frames. Furthermore, SeC adaptively balances LVLM-based semantic reasoning with enhanced feature matching, dynamically adjusting computational efforts based on scene complexity. To rigorously assess VOS methods in scenarios demanding high-level conceptual reasoning and robust semantic understanding, we introduce the Semantic Complex Scenarios Video Object Segmentation benchmark (SeCVOS). SeCVOS comprises 160 manually annotated multi-scenario videos designed to challenge models with substantial appearance variations and dynamic scene transformations. In particular, SeC achieves an 11.8-point improvement over SAM 2.1 on SeCVOS, establishing a new state-of-the-art in concept-aware video object segmentation.

The transformer architecture [1] has fundamentally transformed natural language processing, establishing itself as the dominant paradigm for language modeling and understanding tasks. However, the field's trajectory toward ever-larger models has created significant computational and economic challenges. Contemporary models such as OpenAI's GPT series, Anthropic's Claude, and Google's Gemini have pushed parameter counts into the hundreds of billions, resulting in unprecedented infrastructure costs that increasingly exceed the economic value these models generate in many practical applications. This scaling trajectory has reached a critical inflection point where the marginal benefits of additional parameters diminish rapidly while computational requirements grow exponentially. Despite this economic reality, there has been surprisingly limited systematic exploration of compact, efficient transformer architectures that could deliver comparable performance at sustainable computational costs. The prevailing assumption that model quality scales monotonically with parameter count has created a significant research gap in the sub-billion parameter regime, leaving unexplored the potential for architectural innovation to compensate for reduced scale. In this work, we challenge this scaling paradigm by presenting Supernova, a 650M parameter decoder-only transformer that demonstrates how careful architectural design and tokenization innovation can achieve performance comparable to significantly larger models while maintaining computational efficiency. Our approach is grounded in three fundamental principles: architectural efficiency through modern component integration, superior tokenization design, and dramatic improvements in data efficiency. 1