Papers

This note is a friendly technical check of arXiv:2511.21731v1. I highlight a few places where the manuscript's interpretation of (i) the reported CHSH/Bell-type calculations and (ii) Bose–Einstein (BE) fits to rank-frequency data seems to go beyond what the stated procedures can firmly support. I also point out one internal inconsistency in the "energy-level spacing" analogy. The aim is constructive: to keep the interesting empirical observations, while making clear what they do (and do not) imply about quantum entanglement in the usual Hilbert-space sense, especially when "energy" is defined by rank.

In this report, we present HyperCLOVA X 32B Think, a vision-language model designed with particular emphasis on reasoning within the Korean linguistic and cultural context, as well as agentic ability. HyperCLOVA X 32B Think is pre-trained with a strong focus on reasoning capabilities and subsequently post-trained to support multimodal understanding, enhanced reasoning, agentic behaviors, and alignment with human preferences. Experimental evaluations against comparably sized models demonstrate that our model achieves strong performance on Korean text-to-text and vision-to-text benchmarks, as well as on agent-oriented evaluation tasks. By open-sourcing HyperCLOVA X 32B Think, we aim to support broader adoption and facilitate further research and innovation across both academic and industrial communities.

We present T3C, a train-once, test-time budget-conditioned compression framework that exposes rank and precision as a controllable deployment knob. T3C combines elastic tensor factorization (maintained up to a maximal rank) with rank-tied mixed-precision quantization and a lightweight controller that maps a latency/energy/size budget token to per-layer rank/bit assignments; the policy snaps to hardware-aligned profiles and is monotone in the budget. A fast, layerwise consistency certificate, computed from spectral proxies and activation statistics, upper-bounds logit drift and regularizes training, yielding a practical reliability signal with negligible overhead. On ImageNet-1k, T3C shifts the vision Pareto frontier: for ResNet-50 at matched accuracy (\leq 0.5% drop), p50 latency is 1.18ms with a 38MB model, outperforming PTQ-8b (1.44ms, 88MB); for ViT-B/16, T3C reaches 2.30ms p50 with 59MB, improving over strong PTQ/QAT baselines. A single T3C checkpoint therefore provides predictable, certificate-backed accuracy-latency-size trade-offs on demand across devices.

Detecting distributional drift in high-dimensional data streams presents fundamental challenges: global comparison methods scale poorly, projection-based approaches lose geometric structure, and re-clustering methods suffer from identity instability. This paper introduces Argus, A framework that reconceptualizes drift detection as tracking local statistics over a fixed spatial partition of the data manifold. The key contributions are fourfold. First, it is proved that Voronoi tessellations over canonical orthonormal frames yield drift metrics that are invariant to orthogonal transformations. The rotations and reflections that preserve Euclidean geometry. Second, it is established that this framework achieves O(N) complexity per snapshot while providing cell-level spatial localization of distributional change. Third, a graph-theoretic characterization of drift propagation is developed that distinguishes coherent distributional shifts from isolated perturbations. Fourth, product quantization tessellation is introduced for scaling to very high dimensions (d>500) by decomposing the space into independent subspaces and aggregating drift signals across subspaces. This paper formalizes the theoretical foundations, proves invariance properties, and presents experimental validation demonstrating that the framework correctly identifies drift under coordinate rotation while existing methods produce false positives. The tessellated approach offers a principled geometric foundation for distribution monitoring that preserves high-dimensional structure without the computational burden of pairwise comparisons.

Graph structures are increasingly used in dialog memory systems, but empirical findings on their effectiveness remain inconsistent, making it unclear which design choices truly matter. We present an experimental, system-oriented analysis of long-term dialog memory architectures. We introduce a unified framework that decomposes dialog memory systems into core components and supports both graph-based and non-graph approaches. Under this framework, we conduct controlled, stage-wise experiments on LongMemEval and HaluMem, comparing common design choices in memory representation, organization, maintenance, and retrieval. Our results show that many performance differences are driven by foundational system settings rather than specific architectural innovations. Based on these findings, we identify stable and reliable strong baselines for future dialog memory research. Code are available at https://github.com/AvatarMemory/UnifiedMem

When competing sellers delegate pricing to a shared AI model, such as a large language model, correlated recommendations combined with performance-driven updates aggregating seller feedback raise a key question: can standard AI deployment practices inadvertently produce supracompetitive pricing? We develop a stylized duopoly model in which two sellers receive pricing recommendations from a shared AI characterized by two parameters: a propensity parameter capturing the model's tendency to set high prices and an output-fidelity parameter measuring alignment between this tendency and actual outputs, with propensity updated via periodic retraining on observed outcomes. We find that configuring AI models for robustness and reproducibility can lead to supracompetitive pricing via a phase transition. Below a critical output-fidelity threshold, competitive pricing is the unique stable outcome. Above it, the model exhibits bistability: both competitive and supracompetitive pricing are locally stable, with the realized outcome determined by the model's initial propensity. Supracompetitive pricing raises average prices, but occasional low-price recommendations complicate detection. With perfect output fidelity, full price coordination emerges from any interior initial propensity. For finite training batches of size $b$, when the initial propensity lies in the supracompetitive basin, the probability of supracompetitive pricing approaches 1 as $b$ increases, with the region of indeterminate outcomes shrinking at rate $O(1/\sqrt{b})$. Any factor reducing alignment between the model's propensity and sellers' actual pricing, whether through diversifying AI providers, introducing recommendation noise, or reducing seller adherence, pushes the market toward competitive outcomes.

Large Language Models (LLMs) have demonstrated strong capabilities in interpreting lengthy, complex legal and policy language. However, their reliability can be undermined by hallucinations and inconsistencies, particularly when analyzing subjective and nuanced documents. These challenges are especially critical in medical coverage policy review, where human experts must be able to rely on accurate information. In this paper, we present an approach designed to support human reviewers by making policy interpretation more efficient and interpretable. We introduce a methodology that pairs a coverage-aware retriever with symbolic rule-based reasoning to surface relevant policy language, organize it into explicit facts and rules, and generate auditable rationales. This hybrid system minimizes the number of LLM inferences required which reduces overall model cost. Notably, our approach achieves a 44% reduction in inference cost alongside a 4.5% improvement in F1 score, demonstrating both efficiency and effectiveness.

The deployment of large language models (LLMs) in real-world clinical applications is constrained by the fundamental trade-off between computational cost and the efficiency of linear-time models. To address this, we propose an LLM-based MambaFormer hybrid Mixture-of-Experts (MoE) framework for efficient medical question-answering (QA) and clinical assistance. The MambaFormer employs a lightweight gating mechanism that performs token-level dynamic routing to a customized Transformer expert (ET5) for short, complex queries or to a State Space Model expert (EMamba) for long, high-throughput sequences. The customized EMamba and ET5 models are tailored to accommodate input sequence dimensionality, embedding structure, sequence length, and target-specific output heads, and are fine-tuned through transfer learning on a new, custom-designed DentalQA dataset. Moreover, intelligent routing decisions are driven by the contextual complexity of token embeddings, normalized sequence length, and domain-aware features, thereby enforcing a Pareto-optimal trade-off between inference latency and prediction accuracy. Furthermore, a novel utility-guided multi-objective loss jointly optimizes decisions, router parameters, routing behavior, expert utilization, and computational cost by adaptively regulating token-level expert activation. Finally, the proposed MambaFormer is cross-validated (holdout) for medical QA on the new, custom-designed DentalQA and PubMedQA datasets and compared with state-of-the-art techniques. The proposed MambaFormer outperforms (BERTScore = 0.9180) with ultra-low latency (0.077 s), delivering a 24.4 speedup over T5-Large and establishing a scalable solution for resource-constrained clinical deployment.

Cloud storage has become the backbone of modern data infrastructure, yet privacy and efficient data retrieval remain significant challenges. Traditional privacy-preserving approaches primarily focus on enhancing database security but fail to address the automatic identification of sensitive information before encryption. This can dramatically reduce query processing time and mitigate errors during manual identification of sensitive information, thereby reducing potential privacy risks. To address this limitation, this research proposes an intelligent privacy-preserving query optimization framework that integrates Named Entity Recognition (NER) to detect sensitive information in queries, utilizing secure data encryption and query optimization techniques for both sensitive and non-sensitive data in parallel, thereby enabling efficient database optimization. Combined deep learning algorithms and transformer-based models to detect and classify sensitive entities with high precision, and the Advanced Encryption Standard (AES) algorithm to encrypt, with blind indexing to secure search functionality of the sensitive data, whereas non-sensitive data was divided into groups using the K-means algorithm, along with a rank search for optimization. Among all NER models, the Deep Belief Network combined with Long Short-Term Memory (DBN-LSTM) delivers the best performance, with an accuracy of 93% and precision (94%), recall, and F1 score of 93%, and 93%, respectively. Besides, encrypted search achieved considerably faster results with the help of blind indexing, and non-sensitive data fetching also outperformed traditional clustering-based searches. By integrating sensitive data detection, encryption, and query optimization, this work advances the state of privacy-preserving computation in modern cloud infrastructures.

We present Racka, a lightweight, continually pretrained large language model designed to bridge the resource gap between Hungarian and high-resource languages such as English and German. Racka employs parameter-efficient continual pretraining via Low-Rank Adaptation (LoRA) on a Qwen-3 4B backbone, making the recipe practical on A100 (40GB)-based HPC clusters with low inter-node bandwidth. To better match the training distribution, we replace and adapt the tokenizer, achieving substantially improved tokenization fertility for Hungarian while maintaining competitive performance in English and German. The model is trained on 160B subword tokens drawn from a mixture of internet and high-quality curated sources, with a composition of 44% Hungarian, 24% English, 21% German, and 11% code. This data mix is chosen to mitigate catastrophic forgetting and preserve high-resource language capabilities during continual pretraining. Our preliminary results indicate modest but stable results in language adaptation.

This work addresses critical challenges to academic integrity, including plagiarism, fabrication, and verification of authorship of educational content, by proposing a Natural Language Processing (NLP)-based framework for authenticating students' content through author attribution and style change detection. Despite some initial efforts, several aspects of the topic are yet to be explored. In contrast to existing solutions, the paper provides a comprehensive analysis of the topic by targeting four relevant tasks, including (i) classification of human and machine text, (ii) differentiating in single and multi-authored documents, (iii) author change detection within multi-authored documents, and (iv) author recognition in collaboratively produced documents. The solutions proposed for the tasks are evaluated on two datasets generated with Gemini using two different prompts, including a normal and a strict set of instructions. During experiments, some reduction in the performance of the proposed solutions is observed on the dataset generated through the strict prompt, demonstrating the complexities involved in detecting machine-generated text with cleverly crafted prompts. The generated datasets, code, and other relevant materials are made publicly available on GitHub, which are expected to provide a baseline for future research in the domain.

This study evaluates the linguistic and clinical suitability of synthetic electronic health records in mental health. First, we describe the rationale and the methodology for creating the synthetic corpus. Second, we examine expressions of agency, modality, and information flow across four clinical genres (Assessments, Correspondence, Referrals and Care plans) with the aim to understand how LLMs grammatically construct medical authority and patient agency through linguistic choices. While LLMs produce coherent, terminology-appropriate texts that approximate clinical practice, systematic divergences remain, including registerial shifts, insufficient clinical specificity, and inaccuracies in medication use and diagnostic procedures. The results show both the potential and limitations of synthetic corpora for enabling large-scale linguistic research otherwise impossible with genuine patient records.

Qualitative data are widespread in domains such as healthcare, marketing, and bioinformatics, where clustering offers a fundamental tool for pattern discovery. A core difficulty of qualitative-data clustering lies in measuring similarity among attribute values that carry no inherent ordering or distance. To recover such relationships, existing studies typically rely on within-dataset co-occurrence statistics. This statistical route, however, becomes unreliable once the sample size is small, and the semantic context of each value is therefore left underexploited. Motivated by this limitation, this paper proposes BREVE (Balanced Representation via External Value Enrichment), a clustering framework that enriches each qualitative value with extra semantic dimensions drawn from an external knowledge base. That is, every unique value is expanded by a dense embedding that encodes its semantic content. To prevent the original value identity from being diluted by the added dimensions, a lightweight one-hot component is further appended. An adaptive weight, guided by cluster compactness, then determines how strongly the enrichment dimensions enter the final representation. With this design, experiments on eight benchmark datasets yield an average ARI rank of 1.3 against seven representative competitors.

Large language models (LLMs) frequently produce inaccurate or fabricated information, known as "hallucinations," which compromises their reliability. Existing approaches often train an "Evil LLM" to deliberately generate hallucinations on curated datasets, using these induced hallucinations to guide contrastive decoding against a reliable "positive model" for hallucination mitigation. However, this strategy is limited by the narrow diversity of hallucinations induced, as Evil LLMs trained on specific error types tend to reproduce only these particular patterns, thereby restricting their overall effectiveness. To address these limitations, we propose DHI (Diverse Hallucination Induction), a novel training framework that enables the Evil LLM to generate a broader range of hallucination types without relying on pre-annotated hallucination data. DHI employs a modified loss function that down-weights the generation of specific factually correct tokens, encouraging the Evil LLM to produce diverse hallucinations at targeted positions while maintaining overall factual content. Additionally, we introduce a causal attention masking adaptation to reduce the impact of this penalization on the generation of subsequent tokens. During inference, we apply an adaptive rationality constraint that restricts contrastive decoding to tokens where the positive model exhibits high confidence, thereby avoiding unnecessary penalties on factually correct tokens. Extensive empirical results show that DHI achieves significant performance gains over other contrastive decoding-based approaches across multiple hallucination benchmarks.

Large language models have achieved significant success in various domains, yet their understanding of lyric-centric knowledge has not been fully explored. In this work, we first introduce PlaylistSense, a dataset to evaluate the playlist understanding capability of language models. PlaylistSense encompasses ten types of user queries derived from common real-world perspectives, challenging LLMs to accurately grasp playlist features and address diverse user intents. Comprehensive evaluations indicate that current general-purpose LLMs still have potential for improvement in playlist understanding. Inspired by this, we introduce SongSage, a large musical language model equipped with diverse lyric-centric intelligence through lyric generative pretraining. SongSage undergoes continual pretraining on LyricBank, a carefully curated corpus of 5.48 billion tokens focused on lyrical content, followed by fine-tuning with LyricBank-SFT, a meticulously crafted instruction set comprising 775k samples across nine core lyric-centric tasks. Experimental results demonstrate that SongSage exhibits a strong understanding of lyric-centric knowledge, excels in rewriting user queries for zero-shot playlist recommendations, generates and continues lyrics effectively, and performs proficiently across seven additional capabilities. Beyond its lyric-centric expertise, SongSage also retains general knowledge comprehension and achieves a competitive MMLU score. We will keep the datasets inaccessible due to copyright restrictions and release the SongSage and training script to ensure reproducibility and support music AI research and applications, the datasets release plan details are provided in the appendix.

This paper introduces KOS-TL (Knowledge Operation System Type Logic), a novel constructive framework designed to provide a rigorous logical foundation for autonomous and executable knowledge systems. Traditional knowledge representation models often suffer from a gap between static symbolic logic and dynamic system execution. To bridge this divide, KOS-TL leverages Dependent Type Theory to unify data, logic, and proof into a singular computational substrate.The architecture of KOS-TL is organized into three hierarchical layers: the Core Layer, which defines the static type universe and constructive primitives; the Kernel Layer, which governs state evolution through an event-driven mechanism characterized by the triple $\langle \Sigma, \textsf{Ev}, \Delta \rangle$; and the Runtime Layer, responsible for the bidirectional refinement of physical signals into logical evidence. We formally define the operational semantics of the system and prove key meta-theoretical properties, including Progress and Evolutionary Consistency, ensuring that the system remains logically self-consistent and free from stuck states during continuous state transitions.By integrating Davidsonian event semantics with Martin-Löf type theory, KOS-TL enables the construction of "proof-carrying knowledge," where every state change in the knowledge base is accompanied by a formal witness of its validity. We demonstrate the practical utility of this logic through application examples in industrial traceability and cross-border financial compliance. Our results suggest that KOS-TL provides a robust, formally verifiable basis for the next generation of intelligent, autonomous operating systems.

Large Language Models (LLMs)-powered code review automation has the potential to transform code review workflows. Despite the advances of LLM-powered code review comment generation approaches, several practical challenges remain for designing enterprise-grade code review automation tools. In particular, this paper aims at answering the practical question: how can we design a review-guided, context-aware, quality-checked code review comment generation without fine-tuning? In this paper, we present RovoDev Code Reviewer, an enterprise-grade LLM-based code review automation tool designed and deployed at scale within Atlassian's development ecosystem with seamless integration into Atlassian's Bitbucket. Through the offline, online, user feedback evaluations over a one-year period, we conclude that RovoDev Code Reviewer is effective in generating code review comments that could lead to code resolution for 38.70% (i.e., comments that triggered code changes in the subsequent commits); and offers the promise of accelerating feedback cycles (i.e., decreasing the PR cycle time by 30.8%), alleviating reviewer workload (i.e., reducing the number of human-written comments by 35.6%), and improving overall software quality (i.e., finding errors with actionable suggestions).

We present RoboPhD, a system where AI agents autonomously conduct research to improve Text-to-SQL performance. RoboPhD implements a closed-loop evolution cycle with two coordinated components: a SQL Generation agent composed of a database analysis script and SQL generation instructions, and an Evolution agent that designs new versions based on performance feedback. Central to the framework is an ELO-based selection mechanism enabling survival-of-the-fittest dynamics while handling non-transitivity in performance. Starting from a naive 70-line baseline, RoboPhD evolves agents through iterative cross-pollination, discovering effective techniques without any external guidance on the Text-to-SQL domain. Our best agent, evolved to 1500 lines over 18 iterations, autonomously discovered strategies such as size-adaptive database analysis that adjusts depth based on schema complexity and SQL generation patterns for column selection, evidence interpretation, and aggregation. Evolution provides the largest gains on cheaper models: while we improve by 2.3 points over a strong Claude Opus 4.5 naive baseline, we show an improvement of 8.9 points over the weaker Claude Haiku model. This enables 'skip a tier' deployment: evolved Haiku exceeds naive Sonnet accuracy, and evolved Sonnet exceeds naive Opus, both at lower cost. The full system achieves 73.67% accuracy on the BIRD test set, demonstrating that AI can autonomously build a strong agentic system with only a trivial human-provided starting point.

End-to-End Speech Translation often shows slower convergence and worse performance when target transcriptions exhibit high variance and semantic ambiguity. We propose Listen, Attend, Understand (LAU), a semantic regularization technique that constrains the acoustic encoder's latent space during training. By leveraging frozen text embeddings to provide a directional auxiliary loss, LAU injects linguistic groundedness into the acoustic representation without increasing inference cost. We evaluate our method on a Bambara-to-French dataset with 30 hours of Bambara speech translated by non-professionals. Experimental results demonstrate that LAU models achieve comparable performance by standard metrics compared to an E2E-ST system pretrained with 100% more data and while performing better in preserving semantic meaning. Furthermore, we introduce Total Parameter Drift as a metric to quantify the structural impact of regularization to demonstrate that semantic constraints actively reorganize the encoder's weights to prioritize meaning over literal phonetics. Our findings suggest that LAU is a robust alternative to post-hoc rescoring and a valuable addition to E2E-ST training, especially when training data is scarce and/or noisy.

We introduce EmoLoom-2B, a lightweight and reproducible pipeline that turns small language models under 2B parameters into fast screening candidates for joint emotion classification and Valence-Arousal-Dominance prediction. To ensure protocol-faithful and fair evaluation, we unify data loading, training, and inference under a single JSON input-output contract and remove avoidable variance by adopting KV-off decoding as the default setting. We incorporate two orthogonal semantic regularizers: a VAD-preserving constraint that aligns generated text with target VAD triples, and a lightweight external appraisal classifier that provides training-time guidance on goal attainment, controllability, certainty, and fairness without injecting long rationales. To improve polarity sensitivity, we introduce Valence Flip augmentation based on mirrored emotional pairs. During supervised fine-tuning, we apply A/B mixture sampling with entropy-aware temperature scheduling to balance coverage and convergence. Using Qwen-1.8B-Chat as the base model, EmoLoom-2B achieves strong performance on GoEmotions and EmpatheticDialogues, and demonstrates robust cross-corpus generalization on DailyDialog. The proposed recipe is budget-aware, auditable, and re-entrant, serving as a dependable screening pass before heavier training or multimodal fusion.

Large Language Models (LLMs) demonstrate remarkable fluency across high-resource languages yet consistently fail to generate coherent text in Kashmiri, a language spoken by approximately seven million people. This performance disparity stems not from inherent model limitations but from a critical scarcity of high-quality training data. Decades of Kashmiri literature remain inaccessible to modern NLP pipelines due to their encoding in the proprietary InPage desktop publishing format. This paper introduces KS-LIT-3M, a curated corpus of 3.1 million words (16.4 million characters) specifically designed for pretraining language models on Kashmiri. The dataset is structured as a single continuous linear text stream, optimized for causal language model training where models learn to predict subsequent tokens from preceding context. The corpus was constructed through the development of a specialized InPage-to-Unicode converter, followed by rigorous preprocessing including English contamination removal, character normalization, and quality validation. Encompassing 131,607 unique words drawn from diverse genres including literary works, journalistic writing, academic texts, and religious scholarship, KS-LIT-3M addresses a fundamental resource gap for Kashmiri language technology. The dataset is released under the CC-BY-4.0 license to facilitate research in Kashmiri natural language processing.

This technical report presents the 600K-KS-OCR Dataset, a large-scale synthetic corpus comprising approximately 602,000 word-level segmented images designed for training and evaluating optical character recognition systems targeting Kashmiri script. The dataset addresses a critical resource gap for Kashmiri, an endangered Dardic language utilizing a modified Perso-Arabic writing system spoken by approximately seven million people. Each image is rendered at 256x64 pixels with corresponding ground-truth transcriptions provided in multiple formats compatible with CRNN, TrOCR, and generalpurpose machine learning pipelines. The generation methodology incorporates three traditional Kashmiri typefaces, comprehensive data augmentation simulating real-world document degradation, and diverse background textures to enhance model robustness. The dataset is distributed across ten partitioned archives totaling approximately 10.6 GB and is released under the CC-BY-4.0 license to facilitate research in low-resource language optical character recognition.

Unsupervised Text Style Transfer (UTST) aims to build a system to transfer the stylistic properties of a given text without parallel text pairs. Compared with text transfer between style polarities, UTST for controllable intensity is more challenging due to the subtle differences in stylistic features across different intensity levels. Faced with the challenges posed by the lack of parallel data and the indistinguishability between adjacent intensity levels, we propose a SFT-then-PPO paradigm to fine-tune an LLM. We first fine-tune the LLM with synthesized parallel data. Then, we further train the LLM with PPO, where the rewards are elaborately designed for distinguishing the stylistic intensity in hierarchical levels. Both the global and local stylistic features are considered to formulate the reward functions. The experiments on two UTST benchmarks showcase that both rewards have their advantages and applying them to LLM fine-tuning can effectively improve the performance of an LLM backbone based on various evaluation metrics. Even for close levels of intensity, we can still observe the noticeable stylistic difference between the generated text.

While LLMs are powerful embedding backbones, their application in training-free settings faces two structural challenges: causal attention restricts early tokens from accessing subsequent context, and the next-token prediction objective biases representations toward generation rather than semantic compression. To address these limitations, we propose KV-Embedding, a framework that activates the latent representation power of frozen LLMs. Our method leverages the observation that the key-value (KV) states of the final token at each layer encode a compressed view of the sequence. By re-routing these states as a prepended prefix, we enable all tokens to access sequence-level context within a single forward pass. To ensure model-agnostic applicability, we introduce an automated layer selection strategy based on intrinsic dimensionality. Evaluations on MTEB across Qwen, Mistral, and Llama backbones show that KV-Embedding outperforms existing training-free baselines by up to 10%, while maintaining robust performance on sequences up to 4,096 tokens. These results demonstrate that internal state manipulation offers an efficient alternative to input modification, and we hope this work encourages further exploration of LLM internals for representation learning.

Small language models (SLMs) offer significant deployment advantages but often struggle to match the dialogue quality of larger models in open-domain settings. In this paper, we propose a multi-dimensional prompt-chaining framework that integrates Naturalness, Coherence, and Engagingness dimensions to enhance human-likeness in open-domain dialogue generation. We apply the framework to two SLMs, TinyLlama and Llama-2-7B, and benchmark their performance against responses generated by substantially larger models, including Llama-2-70B and GPT-3.5 Turbo. We then employ automatic and human evaluation to assess the responses based on diversity, contextual coherence, as well as overall quality. Results show that the full framework improves response diversity by up to 29%, contextual coherence by up to 28%, and engagingness as well as naturalness by up to 29%. Notably, Llama-2-7B achieves performance comparable to substantially larger models, including Llama-2-70B and GPT-3.5 Turbo. Overall, the findings demonstrate that carefully designed prompt-based strategies provide an effective and resource-efficient pathway to improving open-domain dialogue quality in SLMs.

From movie characters to modern science fiction - bringing characters into interactive, story-driven conversations has captured imaginations across generations. Achieving this vision is highly challenging and requires much more than just language modeling. It involves numerous complex AI challenges, such as conversational AI, maintaining character integrity, managing personality and emotions, handling knowledge and memory, synthesizing voice, generating animations, enabling real-world interactions, and integration with physical environments. Recent advancements in the development of foundation models, prompt engineering, and fine-tuning for downstream tasks have enabled researchers to address these individual challenges. However, combining these technologies for interactive characters remains an open problem. We present a system and platform for conveniently designing believable digital characters, enabling a conversational and story-driven experience while providing solutions to all of the technical challenges. As a proof-of-concept, we introduce Digital Einstein, which allows users to engage in conversations with a digital representation of Albert Einstein about his life, research, and persona. While Digital Einstein exemplifies our methods for a specific character, our system is flexible and generalizes to any story-driven or conversational character. By unifying these diverse AI components into a single, easy-to-adapt platform, our work paves the way for immersive character experiences, turning the dream of lifelike, story-based interactions into a reality.

As a pivotal task in data lake management, joinable table discovery has attracted widespread interest. While existing language model-based methods achieve remarkable performance by combining offline column representation learning with online ranking, their design insufficiently accounts for the underlying structural interactions: (1) offline, they directly model tables into isolated or pairwise columns, thereby struggling to capture the rich inter-table and intra-table structural information; and (2) online, they rank candidate columns based solely on query-candidate similarity, ignoring the mutual interactions among the candidates, leading to incoherent result sets. To address these limitations, we propose HyperJoin, a large language model (LLM)-augmented Hypergraph framework for Joinable table discovery. Specifically, we first construct a hypergraph to model tables using both the intra-table hyperedges and the LLM-augmented inter-table hyperedges. Consequently, the task of joinable table discovery is formulated as link prediction on this constructed hypergraph. We then design HIN, a Hierarchical Interaction Network that learns expressive column representations through bidirectional message passing over columns and hyperedges. To strengthen coherence and internal consistency in the result columns, we cast online ranking as a coherence-aware top-k column selection problem. We then introduce a reranking module that leverages a maximum spanning tree algorithm to prune noisy connections and maximize coherence. Experiments demonstrate the superiority of HyperJoin, achieving average improvements of 21.4% (Precision@15) and 17.2% (Recall@15) over the best baseline.

Language generation maps a rich, high-dimensional internal state to a single token sequence. We study this many-to-one mapping through the lens of intention collapse: the projection from an internal intention space I to an external language space L. We introduce three cheap, model-agnostic metrics computed on a pre-collapse state I: (i) intention entropy Hint(I), (ii) effective dimensionality deff(I), and (iii) recoverability Recov(I), operationalized as probe AUROC for predicting eventual success. We evaluate these metrics in a 3x3 study across models (Mistral-7B, LLaMA-3.1-8B, Qwen-2.5-7B) and benchmarks (GSM8K, ARC-Challenge, AQUA-RAT), comparing baseline, chain-of-thought (CoT), and a babble control (n=200 items per cell). CoT increases average accuracy from 34.2% to 47.3% (+13.1 pp), driven by large gains on GSM8K but consistent degradations on ARC-Challenge. Across models, CoT induces distinct entropy regimes relative to baseline, dH = Hint(CoT) - Hint(Base): Mistral shows dH < 0 (lower-entropy CoT), whereas LLaMA shows dH > 0 (higher-entropy CoT), highlighting heterogeneity in CoT-induced internal uncertainty. Finally, probe AUROC is significantly above chance in a subset of settings and can dissociate from behavioral accuracy (e.g., high AUROC alongside lower CoT accuracy on ARC-Challenge for Qwen), suggesting that informative internal signal is not always reliably converted into a final discrete decision under constrained response formats.

AI agents increasingly assist with financial research, yet no benchmark evaluates their ability to retrieve specific numeric values from structured databases. We introduce FinRetrieval, a benchmark of 500 financial retrieval questions with ground truth answers, agent responses from 14 configurations across three frontier providers (Anthropic, OpenAI, Google), and complete tool call execution traces. Our evaluation reveals that tool availability dominates performance: Claude Opus achieves 90.8% accuracy with structured data APIs but only 19.8% with web search alone–a 71 percentage point gap that exceeds other providers by 3-4x. We find that reasoning mode benefits vary inversely with base capability (+9.0pp for OpenAI vs +2.8pp for Claude), explained by differences in base-mode tool utilization rather than reasoning ability. Geographic performance gaps (5.6pp US advantage) stem from fiscal year naming conventions, not model limitations. We release the dataset, evaluation code, and tool traces to enable research on financial AI systems.

The rapid growth of Retrieval-Augmented Generation (RAG) has created a proliferation of toolkits, yet a fundamental gap remains between experimental prototypes and robust, production-ready systems. We present SearchGym, a modular infrastructure designed for cross-platform benchmarking and hybrid search orchestration. Unlike existing model-centric frameworks, SearchGym decouples data representation, embedding strategies, and retrieval logic into stateful abstractions: Dataset, VectorSet, and App. This separation enables a Compositional Config Algebra, allowing designers to synthesize entire systems from hierarchical configurations while ensuring perfect reproducibility. Moreover, we analyze the "Top-$k$ Cognizance" in hybrid retrieval pipelines, demonstrating that the optimal sequence of semantic ranking and structured filtering is highly dependent on filter strength. Evaluated on the LitSearch expert-annotated benchmark, SearchGym achieves a 70% Top-100 retrieval rate. SearchGym reveals a design tension between generalizability and optimizability, presenting the potential where engineering optimization may serve as a tool for uncovering the causal mechanisms inherent in information retrieval across heterogeneous domains. An open-source implementation of SearchGym is available at: https://github.com/JeromeTH/search-gym