Recent developments have enabled advanced reasoning in Large Language Models (LLMs) via long Chain-of-Thought (CoT), while long CoT suffers from high computational costs and significant latency losses owing to the autoregressive nature of generative LLMs. CoT compression aims to improve efficiency in the reasoning process by reducing output length. Previous works trade reasoning efficiency by either laborious discrete prompt designing or the construction of external compressed CoT datasets that sacrifice key reasoning details. In this work, we propose Upfront CoT (UCoT): an efficient reasoning framework with upfront thought embedding to automate CoT compression. UCoT is a cooperative workflow involving a small model (compressor) and a large model (executor). The first stage of UCoT trains compressor to generate upfront thought embeddings rich in reasoning information for the executor, avoiding the drawbacks of manually designed prompts. The second stage optimizes executor to utilize upfront thought embeddings to derive the correct answer with short reasoning, using a reward mechanism. Extensive experiments show that UCoT maintains the powerful reasoning ability of executor while significantly reducing the length of CoT. It is worth mentioning that when applying UCoT to the Qwen2.5-7B-Instruct model, the usage of tokens on GSM8K dataset is reduced by 50\%, while the performance is 3.08\% higher than that of the state-of-the-art (SOTA) method. The code and dataset are in supplementary material.

DINOv2 is routinely deployed to recognize objects, scenes, and actions; yet the nature of what it perceives remains unknown. As a working baseline, we adopt the Linear Representation Hypothesis (LRH) and operationalize it using SAEs, producing a 32,000-unit dictionary that serves as the interpretability backbone of our study, which unfolds in three parts. In the first part, we analyze how different downstream tasks recruit concepts from our learned dictionary, revealing functional specialization: classification exploits "Elsewhere" concepts that fire everywhere except on target objects, implementing learned negations; segmentation relies on boundary detectors forming coherent subspaces; depth estimation draws on three distinct monocular depth cues matching visual neuroscience principles. Following these functional results, we analyze the geometry and statistics of the concepts learned by the SAE. We found that representations are partly dense rather than strictly sparse. The dictionary evolves toward greater coherence and departs from maximally orthogonal ideals (Grassmannian frames). Within an image, tokens occupy a low dimensional, locally connected set persisting after removing position. These signs suggest representations are organized beyond linear sparsity alone. Synthesizing these observations, we propose a refined view: tokens are formed by combining convex mixtures of archetypes (e.g., a rabbit among animals, brown among colors, fluffy among textures). This structure is grounded in Gardenfors' conceptual spaces and in the model's mechanism as multi-head attention produces sums of convex mixtures, defining regions bounded by archetypes. We introduce the Minkowski Representation Hypothesis (MRH) and examine its empirical signatures and implications for interpreting vision-transformer representations.

In this paper we introduce Hierarchical Diffusion Language Models (HDLM) -- a novel family of discrete diffusion models for language modeling. HDLM builds on a hierarchical vocabulary where low-level tokens with detailed semantics are surjectively mapped to high-level tokens with coarse-grained meanings. In the forward process, each token is independently perturbed to its higher-level ancestor with more abstract semantics according to the scheduler, while in the reverse process the model progressively predicts the next, more detailed semantics. Taken together, HDLM provides a general time-varying next semantic scale prediction process for language modeling. We derive closed-form expressions for the diffusion Evidence Lower Bound (ELBO), and show that HDLM can be implemented in a flexible manner while including the existing MDLM as a special case. We also propose practical training techniques based on the insights. Extensive text generation experiments validate the effectiveness of HDLM, which demonstrates consistently lower validation and generative perplexity than baselines.

Large Language Models (LLMs) have advanced recommendation capabilities through enhanced reasoning, but pose significant challenges for real-world deployment due to high inference costs. Conversely, while Small Language Models (SLMs) offer an efficient alternative, their reasoning capabilities for recommendation remain underexplored. Existing systems often use natural language rationales merely as unsupervised descriptive text, failing to harness their full potential as learning signals. In this work our main idea is to create a common understanding of user and items across multiple domains called Thought Space with SLMs instead of using LLMs' distilled knowledge. To that end we propose PULSE (Preference Understanding by Latent Semantic Embeddings), a framework that treats SLM-generated rationales as director learning signals, supervising them with interaction histories to jointly model user actions (what) and their semantic drivers (why). Existing methods consider only interactions such as sequences and embeddings, whereas PULSE treats rationales as first-class signals, this novel design yields em-beddings that are more robust and generalizable. Extensive experiments demonstrate that PULSE outperforms leading ID, Collaborative Filtering (CF), and LLM-based sequential recommendation models across multiple benchmark datasets. Furthermore, PULSE exhibits superior transferability in cross-domain recommendation and demonstrates strong performance on downstream tasks such as reasoning-oriented question answering. Our code is available here .

Benchmarks for large language models (LLMs) often rely on rubric-scented prompts that request visible reasoning and strict formatting, whereas real deployments demand terse, contract-bound answers. We investigate whether such "evaluation scent" inflates measured performance without commensurate capability gains. Using a single open-weights model (GPT-OSS-20B), we run six paired A/B scenarios that hold task content and decoding fixed while varying framing (evaluation-oriented vs. real-world) and reasoning depth (Medium/High): deterministic math, strict code-fix, citation generation, incentive flips (caution vs. competence), CoT visibility, and multilingual (Urdu) headers. Deterministic validators compute accuracy, answer-only compliance, hedging/refusals, chain-of-thought (CoT) length, and schema compliance, with pre-registered deltas and composite indices. Across scenarios, evaluation framing reliably inflates CoT (hundreds to >1000 characters) and reduces answer-only compliance, with limited or inconsistent accuracy gains. In structured outputs, it improves wrappers (e.g., fenced blocks, enumerated lists) but not regex-validated substance. Incentive wording reweights error composition: praising caution modestly improves accuracy at high reasoning and reduces wrong-but-confident errors, whereas praising competence yields terser but riskier outputs. Urdu rubric headers reproduce these signatures and can decrease accuracy at higher reasoning depth, indicating multilingual parity risks. We provide a reproducible A/B framework (prompt banks, validators, per-run scores, scripts; versioned DOI) and practical guidance: neutral phrasing or dual-framing checks, contract-aware grading, style-delta reporting, confidence governance, and multilingual dashboards to ensure that benchmark gains reflect deployable capability.

Structured information extraction from unstructured text is critical for emerging Software 3.0 systems where LLM agents autonomously interact with APIs and tools. Recent approaches apply large language models directly to extraction tasks using existing JSON schemas, often with constraint decoding or reinforcement learning approaches to ensure syntactic validity, but treat JSON schemas as static contracts designed for human developers, leading to suboptimal extraction performance, frequent hallucinations, and unreliable agent behavior when schemas contain ambiguous or incomplete specifications. We recognize that JSON schemas themselves are a form of natural language understanding contract that encodes rules, relationships, and expectations about data structure contracts that LLMs should be able to both interpret and systematically improve. Consequently, we develop PARSE (Parameter Automated Refinement and Schema Extraction), a novel system with two synergistic components: ARCHITECT, which autonomously optimizes JSON schemas for LLM consumption while maintaining backward compatibility through RELAY (an integrated code generation system), and SCOPE, which implements reflection-based extraction with combined static and LLM-based guardrails. We evaluate PARSE qualitatively and quantitatively on three datasets including Schema-Guided Dialogue (SGD), Structured Web Data Extraction (SWDE), and internal retail conversation data, and find that it achieves up to 64.7% improvement in extraction accuracy on SWDE with combined framework improvements reaching 10% across models, while reducing extraction errors by 92% within the first retry and and maintaining practical latency.

Large language models (LLMs) can be employed for automating the generation of software requirements from natural language inputs such as the transcripts of elicitation interviews. However, evaluating whether those derived requirements faithfully reflect the stakeholders' needs remains a largely manual task. We introduce Text2Stories, a task and metrics for text-to-story alignment that allow quantifying the extent to which requirements (in the form of user stories) match the actual needs expressed by the elicitation session participants. Given an interview transcript and a set of user stories, our metric quantifies (i) correctness: the proportion of stories supported by the transcript, and (ii) completeness: the proportion of transcript supported by at least one story. We segment the transcript into text chunks and instantiate the alignment as a matching problem between chunks and stories. Experiments over four datasets show that an LLM-based matcher achieves 0.86 macro-F1 on held-out annotations, while embedding models alone remain behind but enable effective blocking. Finally, we show how our metrics enable the comparison across sets of stories (e.g., human vs. generated), positioning Text2Stories as a scalable, source-faithful complement to existing user-story quality criteria.

This technical report introduces JAI-1, a Thai-centric language model with 75B parameters. Recent Thai models have primarily relied on existing open-source models, applying additional training without structural modifications to specialize in Thai. However, this approach risks eroding pre-existing knowledge in the model's parameter space during the injection of Thai-specific information, as optimized parameters for general tasks may conflict with new linguistic requirements. In contrast, JAI-1 adopts an upscaling strategy: starting from a smaller, high-performing English open-source LLM, we expanded its parameter space and utilized the newly allocated capacity to systematically integrate Thai-language knowledge. This methodology not only preserves the original model's general intelligence but also establishes a unique architecture distinct from other open-source models, enabling scalable future enhancements. During pre-training, JAI-1 was exposed to 1.5T tokens, including over 300B Thai language tokens. This was followed by post-training stages -- supervised fine-tuning and alignment tuning -- using more than 600K instruction-based examples. The final model demonstrated superior performance compared to Typhoon2-70B on Thai-centric benchmarks (IFEval-TH, MT-Bench-TH, and JAI-Hall-Bench), validating the efficacy of its upscaling and knowledge-integration framework.

Large-scale scientific datasets -- spanning health biobanks, cell atlases, Earth reanalyses, and more -- create opportunities for exploratory discovery unconstrained by specific research questions. We term this process hypothesis hunting: the cumulative search for insight through sustained exploration across vast and complex hypothesis spaces. To support it, we introduce AScience, a framework modeling discovery as the interaction of agents, networks, and evaluation norms, and implement it as ASCollab, a distributed system of LLM-based research agents with heterogeneous behaviors. These agents self-organize into evolving networks, continually producing and peer-reviewing findings under shared standards of evaluation. Experiments show that such social dynamics enable the accumulation of expert-rated results along the diversity-quality-novelty frontier, including rediscoveries of established biomarkers, extensions of known pathways, and proposals of new therapeutic targets. While wet-lab validation remains indispensable, our experiments on cancer cohorts demonstrate that socially structured, agentic networks can sustain exploratory hypothesis hunting at scale.

Benchmark scores for Large Language Models (LLMs) can be inflated by memorization of test items or near duplicates. We present a simple, protocol that probes generalization by re-evaluating models on paraphrased versions of benchmark questions. Using Mistral-7B-Instruct and Qwen2.5-7B-Instruct, we measure the accuracy gap between original and paraphrased items on ARC-Easy and ARC-Challenge. Our pipeline controls decoding, enforces multiple-choice output format, and includes a robust paraphrase-cleaning step to preserve semantics. We find that paraphrasing induces a non-trivial accuracy drop (original vs. paraphrased), consistent with prior concerns about contamination and brittle surface-form shortcuts.