Daily Papers

Self-consistency detects hallucinations by generating multiple sampled answers to a question and measuring agreement, but this requires repeated decoding and can be sensitive to lexical variation. Semantic self-consistency improves this by clustering sampled answers by meaning using natural language inference, but it adds both sampling cost and external inference overhead. We show that first-token confidence, phi_first, computed from the normalized entropy of the top-K logits at the first content-bearing answer token of a single greedy decode, matches or modestly exceeds semantic self-consistency on closed-book short-answer factual question answering. Across three 7-8B instruction-tuned models and two benchmarks, phi_first achieves a mean AUROC of 0.820, compared with 0.793 for semantic agreement and 0.791 for standard surface-form self-consistency. A subsumption test shows that phi_first is moderately to strongly correlated with semantic agreement, and combining the two signals yields only a small AUROC improvement over phi_first alone. These results suggest that much of the uncertainty information captured by multi-sample agreement is already available in the model's initial token distribution. We argue that phi_first should be reported as a default low-cost baseline before invoking sampling-based uncertainty estimation.

In the rapidly evolving semiconductor industry, where research, design, verification, and manufacturing are intricately linked, the potential of Large Language Models to revolutionize hardware design and security verification is immense. The primary challenge, however, lies in the complexity of hardware specific issues that are not adequately addressed by the natural language or software code knowledge typically acquired during the pretraining stage. Additionally, the scarcity of datasets specific to the hardware domain poses a significant hurdle in developing a foundational model. Addressing these challenges, this paper introduces Hardware Phi 1.5B, an innovative large language model specifically tailored for the hardware domain of the semiconductor industry. We have developed a specialized, tiered dataset comprising small, medium, and large subsets and focused our efforts on pretraining using the medium dataset. This approach harnesses the compact yet efficient architecture of the Phi 1.5B model. The creation of this first pretrained, hardware domain specific large language model marks a significant advancement, offering improved performance in hardware design and verification tasks and illustrating a promising path forward for AI applications in the semiconductor sector.

We introduce Phi-4-Mini and Phi-4-Multimodal, compact yet highly capable language and multimodal models. Phi-4-Mini is a 3.8-billion-parameter language model trained on high-quality web and synthetic data, significantly outperforming recent open-source models of similar size and matching the performance of models twice its size on math and coding tasks requiring complex reasoning. This achievement is driven by a carefully curated synthetic data recipe emphasizing high-quality math and coding datasets. Compared to its predecessor, Phi-3.5-Mini, Phi-4-Mini features an expanded vocabulary size of 200K tokens to better support multilingual applications, as well as group query attention for more efficient long-sequence generation. Phi-4-Multimodal is a multimodal model that integrates text, vision, and speech/audio input modalities into a single model. Its novel modality extension approach leverages LoRA adapters and modality-specific routers to allow multiple inference modes combining various modalities without interference. For example, it now ranks first in the OpenASR leaderboard to date, although the LoRA component of the speech/audio modality has just 460 million parameters. Phi-4-Multimodal supports scenarios involving (vision + language), (vision + speech), and (speech/audio) inputs, outperforming larger vision-language and speech-language models on a wide range of tasks. Additionally, we experiment to further train Phi-4-Mini to enhance its reasoning capabilities. Despite its compact 3.8-billion-parameter size, this experimental version achieves reasoning performance on par with or surpassing significantly larger models, including DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B.

Large vision-language models (VLMs) are increasingly deployed for optical character recognition (OCR) in healthcare settings, raising critical concerns about protected health information (PHI) exposure during document processing. This work presents the first systematic evaluation of inference-time vision token masking as a privacy-preserving mechanism for medical document OCR using DeepSeek-OCR. We introduce seven masking strategies (V3-V9) targeting different architectural layers (SAM encoder blocks, compression layers, dual vision encoders, projector fusion) and evaluate PHI reduction across HIPAA-defined categories using 100 synthetic medical billing statements (drawn from a corpus of 38,517 annotated documents) with perfect ground-truth annotations. All masking strategies converge to 42.9% PHI reduction, successfully suppressing long-form spatially-distributed identifiers (patient names, dates of birth, physical addresses at 100% effectiveness) while failing to prevent short structured identifiers (medical record numbers, social security numbers, email addresses, account numbers at 0% effectiveness). Ablation studies varying mask expansion radius (r=1,2,3) demonstrate that increased spatial coverage does not improve reduction beyond this ceiling, indicating that language model contextual inference - not insufficient visual masking - drives structured identifier leakage. A simulated hybrid architecture combining vision masking with NLP post-processing achieves 88.6% total PHI reduction (assuming 80% NLP accuracy on remaining identifiers). This negative result establishes boundaries for vision-only privacy interventions in VLMs, provides guidance distinguishing PHI types amenable to vision-level versus language-level redaction, and redirects future research toward decoder-level fine-tuning and hybrid defense-in-depth architectures for HIPAA-compliant medical document processing.

With the development of Multimodal Large Language Models (MLLMs) technology, its general capabilities are increasingly powerful. To evaluate the various abilities of MLLMs, numerous evaluation systems have emerged. But now there is still a lack of a comprehensive method to evaluate MLLMs in the tasks related to flowcharts, which are very important in daily life and work. We propose the first comprehensive method, FlowCE, to assess MLLMs across various dimensions for tasks related to flowcharts. It encompasses evaluating MLLMs' abilities in Reasoning, Localization Recognition, Information Extraction, Logical Verification, and Summarization on flowcharts. However, we find that even the GPT4o model achieves only a score of 56.63. Among open-source models, Phi-3-Vision obtained the highest score of 49.97. We hope that FlowCE can contribute to future research on MLLMs for tasks based on flowcharts. https://github.com/360AILABNLP/FlowCE

We present DreamX-Phi 1.0, an action-conditioned video world model for robotic manipulation that, given an observed frame, a language instruction, and a prescribed action sequence comprising end-effector poses and gripper states, predicts the resulting future observations. Yet realism alone does not guarantee faithfulness: a convincing rollout can still move the wrong arm or lose the manipulated object. To ensure the prediction respects each arm's commanded path, we inject per-arm SE(3) transformations into attention via PRoPE-style geometric encoding, preserving arm identity and rigid-motion structure. Action control alone does not fully constrain scene geometry or the evolution of small manipulated objects. We therefore add a lightweight depth branch for scene-level geometry and use SAM3 masks with a frozen V-JEPA teacher to maintain object consistency throughout grasping. We further distill the multi-step generator into a few-step student via distribution-matching distillation for efficient deployment. At the time of writing, achieves first place on Track~1 and second place on Track~2 of the WorldArena~2.0 Challenge. Our model and code will be publicly available.

In this paper cum technical report, we present PHUDGE A fine tuned Phi3 model that achieved SOTA results in 4 tasks as Feedback Test, Feedback OOD, MT Human, Preference Test surpassing each and every existing model in latency and throughput. It shows very strong correlation not only with GPT4 but with Human annotators too in unseen data as well as in both absolute and relative grading tasks. We have not only addressed the usage of small LMs for cost effective production grade systems but have also shown that Causal modelling is not only slow in nature but sometimes it can hinder models learning capabilities and should be replaced by simpler tasks whenever we can to make the overall system faster and better. We show that by following systematic ML experimentation, thoughtful data augmentation and re purposing the problem itself, we can even beat 10x bigger models even with lesser training data. To the best of our knowledge, we are re the first one to experiment and showcase the usage of generalised version of Earth Movers Distance AKA Wasserstein distance by using Minkowski Distance with a penalty to control loss smoothing and can be used as a loss function instead of Cross Entropy to get stable training and better results for grading tasks.

We introduce TeMLM, a set of transparency-first release artifacts for clinical language models. TeMLM unifies provenance, data transparency, modeling transparency, and governance into a single, machine-checkable release bundle. We define an artifact suite (TeMLM-Card, TeMLM-Datasheet, TeMLM-Provenance) and a lightweight conformance checklist for repeatable auditing. We instantiate the artifacts on Technetium-I, a large-scale synthetic clinical NLP dataset with 498,000 notes, 7.74M PHI entity annotations across 10 types, and ICD-9-CM diagnosis labels, and report reference results for ProtactiniumBERT (about 100 million parameters) on PHI de-identification (token classification) and top-50 ICD-9 code extraction (multi-label classification). We emphasize that synthetic benchmarks are valuable for tooling and process validation, but models should be validated on real clinical data prior to deployment.

Test-time scaling (TTS), which involves dynamic allocation of compute during inference, offers a promising way to improve reasoning in large language models. While existing TTS methods work well, they often rely on long decoding paths or require a large number of samples to be generated, increasing the token usage and inference latency. We observe the surprising fact that for reasoning tasks, shorter traces are much more likely to be correct than longer ones. Motivated by this, we introduce First Finish Search (FFS), a training-free parallel decoding strategy that launches n independent samples and returns as soon as any one completes. We evaluate FFS alongside simple decoding, beam search, majority voting, and budget forcing on four reasoning models (DeepSeek-R1, R1-Distill-Qwen-32B, QwQ-32B and Phi-4-Reasoning-Plus) and across four datasets (AIME24, AIME25-I, AIME25-II and GPQA Diamond). With DeepSeek-R1, FFS achieves 82.23% accuracy on the AIME datasets, a 15% improvement over DeepSeek-R1's standalone accuracy, nearly matching OpenAI's o4-mini performance. Our theoretical analysis explains why stopping at the shortest trace is likely to yield a correct answer and identifies the conditions under which early stopping may be suboptimal. The elegance and simplicity of FFS demonstrate that straightforward TTS strategies can perform remarkably well, revealing the untapped potential of simple approaches at inference time.

We present a search for the diffuse extremely-high-energy neutrino flux using 12.6 years of IceCube data. The non-observation of neutrinos with energies well above 10 , PeV constrains the all-flavor neutrino flux at 10^{18} , eV to a level of E^2 Phi_{nu_e + nu_mu + nu_tau} simeq 10^{-8} , GeV , cm^{-2} , s^{-1} , sr^{-1}, the most stringent limit to date. Using this data, we constrain the proton fraction of ultra-high-energy cosmic rays (UHECRs) above simeq 30 , EeV to be lesssim 70,% (at 90,% CL) if the cosmological evolution of the sources is comparable to or stronger than the star formation rate. This result complements direct air-shower measurements by being insensitive to uncertainties associated with hadronic interaction models. It is the first such result to disfavor the ``proton-only" hypothesis for UHECRs using neutrino data.

As AI systems become more integrated into daily life, the need for safer and more reliable moderation has never been greater. Large Language Models (LLMs) have demonstrated remarkable capabilities, surpassing earlier models in complexity and performance. Their evaluation across diverse tasks has consistently showcased their potential, enabling the development of adaptive and personalized agents. However, despite these advancements, LLMs remain prone to errors, particularly in areas requiring nuanced moral reasoning. They struggle with detecting implicit hate, offensive language, and gender biases due to the subjective and context-dependent nature of these issues. Moreover, their reliance on training data can inadvertently reinforce societal biases, leading to inconsistencies and ethical concerns in their outputs. To explore the limitations of LLMs in this role, we developed an experimental framework based on state-of-the-art (SOTA) models to assess human emotions and offensive behaviors. The framework introduces a unified benchmark dataset encompassing 49 distinct categories spanning the wide spectrum of human emotions, offensive and hateful text, and gender and racial biases. Furthermore, we introduced SafePhi, a QLoRA fine-tuned version of Phi-4, adapting diverse ethical contexts and outperforming benchmark moderators by achieving a Macro F1 score of 0.89, where OpenAI Moderator and Llama Guard score 0.77 and 0.74, respectively. This research also highlights the critical domains where LLM moderators consistently underperformed, pressing the need to incorporate more heterogeneous and representative data with human-in-the-loop, for better model robustness and explainability.

Most of the current task-oriented dialogue systems (ToD), despite having interesting results, are designed for a handful of languages like Chinese and English. Therefore, their performance in low-resource languages is still a significant problem due to the absence of a standard dataset and evaluation policy. To address this problem, we proposed ViWOZ, a fully-annotated Vietnamese task-oriented dialogue dataset. ViWOZ is the first multi-turn, multi-domain tasked oriented dataset in Vietnamese, a low-resource language. The dataset consists of a total of 5,000 dialogues, including 60,946 fully annotated utterances. Furthermore, we provide a comprehensive benchmark of both modular and end-to-end models in low-resource language scenarios. With those characteristics, the ViWOZ dataset enables future studies on creating a multilingual task-oriented dialogue system.

Training a Mixture-of-Experts (MoE) model at long context or large batch size fails as soon as any one component's peak allocation exceeds device memory, so the target is every peak at once, not the average footprint. Four are left unbounded by the parallelism plans in common use, and each grows differently: expert dispatch with the routing matrix, the vocabulary projection with tokens times vocabulary, gradient checkpoint boundaries with depth times sequence length, and optimizer state with parameter count. Which one runs out first changes with the model, the context length, and the device count, so lowering the largest only exposes the next. We bound all four with schedules whose GPU working set is fixed at launch: PipelinedLLEP extends least-loaded expert parallelism with a cap on the tokens each source contributes to a dispatch chunk, Ring-DTP circulates activations or weight shards around a ring at the vocabulary projection and folds each block of logits into an online log-sum-exp, Selective checkpoint offload (SCO) keeps the one long-lived tensor of each checkpoint boundary in CPU memory, and OffloadStreamAdamW turns the serial CPU Adam update of optimizer offload into a bucket pipeline. All four change only the order and granularity of computation and data movement, so the loss and gradients stay exact. In matched component tests, they cut the MoE dispatch peak by up to 59.3% without losing throughput, the vocabulary projection peak by 86.6%, and the offloaded optimizer step by 2.05times faster. Composed on MoE models from 120B to 667B parameters, they train at 1M context length, 8--32times the reach of a tuned FSDP2 baseline, and up to 10.4times its throughput.

Modern LLM-based agents operate through a harness of tools, reusable skills, and specialist agents that shapes what they observe and what they can do. In practice, this harness continually evolves as new capabilities are added. We introduce EVOHARNESSBENCH, a benchmark for evaluating agents under controlled harness evolution across three axes (tools, skills, and agents). Unlike existing continual-learning benchmarks for agents, which typically place non-stationarity (i.e., what changes over time) in the task stream while keeping the harness fixed, EVOHARNESSBENCH places non-stationarity in the externally supplied harness itself. It contains 17 multi-stage harness streams constructed deterministically from verifier-based benchmarks, comprising 802 tasks, 520 tools, 42 skills, and 62 agents. We evaluate two complementary settings corresponding to the central challenges of harness evolution: deployment evaluation, which isolates retention of previously accessible competence as the harness expands, and self-evolving adaptation evaluation, which tests whether accumulated experience remains useful as new capabilities are introduced. Our results reveal three persistent gaps. First, harness expansion alone can degrade performance on previously solved tasks, producing harness-induced forgetting. Second, gains from self-evolving adaptation remain inconsistent across stages of harness evolution, capability axes, and environments. Third, retention and adaptation can pull in different directions: preserving earlier competence does not necessarily improve adaptation to newly introduced capabilities, and vice versa. These results establish harness evolution as a distinct challenge for building agents that can keep pace with an evolving harness while preserving previously effective behavior.

Multi-agent systems (MAS) leveraging the impressive capabilities of Large Language Models (LLMs) hold significant potential for tackling complex tasks. However, most current MAS depend on manually designed agent roles and communication protocols. These manual designs often fail to align with the underlying LLMs' strengths and struggle to adapt to novel tasks. Recent automatic MAS approaches attempt to mitigate these limitations but typically necessitate a validation set for tuning and yield static MAS designs lacking adaptability during inference. We introduce MAS-ZERO, the first self-evolved, inference-time framework for automatic MAS design. MAS-ZERO employs meta-level design to iteratively generate, evaluate, and refine MAS configurations tailored to each problem instance, without requiring a validation set. Critically, it enables dynamic agent composition and problem decomposition through meta-feedback on solvability and completeness. Experiments across math, graduate-level QA, and software engineering benchmarks, using both closed-source and open-source LLM backbones of varying sizes, demonstrate that MAS-ZERO outperforms both manual and automatic MAS baselines, achieving a 7.44% average accuracy improvement over the next strongest baseline while maintaining cost-efficiency. These findings underscore the promise of meta-level self-evolved design for creating effective and adaptive MAS.

While knowledge distillation has seen widespread use in pre-training and instruction tuning, its application to aligning language models with human preferences remains underexplored, particularly in the more realistic cross-tokenizer setting. The incompatibility of tokenization schemes between teacher and student models has largely prevented fine-grained, white-box distillation of preference information. To address this gap, we propose Cross-Tokenizer Preference Distillation (CTPD), the first unified framework for transferring human-aligned behavior between models with heterogeneous tokenizers. CTPD introduces three key innovations: (1) Aligned Span Projection, which maps teacher and student tokens to shared character-level spans for precise supervision transfer; (2) a cross-tokenizer adaptation of Token-level Importance Sampling (TIS-DPO) for improved credit assignment; and (3) a Teacher-Anchored Reference, allowing the student to directly leverage the teacher's preferences in a DPO-style objective. Our theoretical analysis grounds CTPD in importance sampling, and experiments across multiple benchmarks confirm its effectiveness, with significant performance gains over existing methods. These results establish CTPD as a practical and general solution for preference distillation across diverse tokenization schemes, opening the door to more accessible and efficient alignment of language models.

Large language model (LLM)-based reasoning systems have recently achieved gold medal-level performance in the IMO 2025 competition, writing mathematical proofs where, to receive full credit, each step must be not only correct but also sufficiently supported. To train LLM-based reasoners in such challenging, open-ended settings, strong verifiers capable of catching step-level mistakes are necessary prerequisites. We introduce Hard2Verify, a human-annotated, step-level verification benchmark produced with over 500 hours of human labor. Hard2Verify is designed to rigorously assess step-level verifiers at the frontier: Verifiers must provide step-level annotations or identify the first error in responses generated by frontier LLMs for very recent, challenging, and open-ended math questions. We evaluate 29 generative critics and process reward models, demonstrating that, beyond a few standouts, open-source verifiers lag closed source models. We subsequently analyze what drives poor performance in step-level verification, the impacts of scaling verifier compute, as well as fundamental questions such as self-verification and verification-generation dynamics.

Large vision-language models (VLMs) demonstrate strong performance in medical image understanding, but frequently generate clinically plausible yet incorrect statements, raising significant safety concerns. Existing medical hallucination benchmarks primarily focus on 2D imaging with one-shot diagnostic questions, offering limited insight into whether predictions are grounded in correct localization and abnormality identification, allowing critical reasoning errors to remain hidden behind seemingly correct diagnoses. We introduce Med-StepBench, the first large-scale benchmark for step-wise hallucination detection in 3D oncological PET/CT, comprising over 12,000 images and more than 1,000,000 image-statement pairs across volumetric and multi-view 2D data, which decomposes clinical reasoning into four expert-designed diagnostic stages. Using clinician-verified annotations, we perform the first step-level evaluation of general-purpose and medical VLMs, revealing systematic failure modes obscured by aggregate accuracy metrics. Furthermore, we show that current VLMs are highly susceptible to adversarial yet clinically plausible intermediate explanations, which significantly amplify hallucinations despite contradictory visual evidence. Together, our findings highlight fundamental limitations in grounding multi-step clinical reasoning and establish Med-StepBench as a rigorous benchmark for developing safer and more reliable medical VLMs.

Understanding the inner workings of neural networks is essential for enhancing model performance and interpretability. Current research predominantly focuses on examining the connection between individual neurons and the model's final predictions. Which suffers from challenges in interpreting the internal workings of the model, particularly when neurons encode multiple unrelated features. In this paper, we propose a novel framework that transitions the focus from analyzing individual neurons to investigating groups of neurons, shifting the emphasis from neuron-output relationships to functional interaction between neurons. Our automated framework, NeurFlow, first identifies core neurons and clusters them into groups based on shared functional relationships, enabling a more coherent and interpretable view of the network's internal processes. This approach facilitates the construction of a hierarchical circuit representing neuron interactions across layers, thus improving interpretability while reducing computational costs. Our extensive empirical studies validate the fidelity of our proposed NeurFlow. Additionally, we showcase its utility in practical applications such as image debugging and automatic concept labeling, thereby highlighting its potential to advance the field of neural network explainability.

The rapid advancement of large language models (LLMs) has enabled new possibilities for applying artificial intelligence within the legal domain. Nonetheless, the complexity, hierarchical organization, and frequent revisions of Vietnamese legislation pose considerable challenges for evaluating how well these models interpret and utilize legal knowledge. To address this gap, the Vietnamese Legal Benchmark (VLegal-Bench) is introduced, the first comprehensive benchmark designed to systematically assess LLMs on Vietnamese legal tasks. Informed by Bloom's cognitive taxonomy, VLegal-Bench encompasses multiple levels of legal understanding through tasks designed to reflect practical usage scenarios. The benchmark comprises 10,450 samples generated through a rigorous annotation pipeline, where legal experts label and cross-validate each instance using our annotation system to ensure every sample is grounded in authoritative legal documents and mirrors real-world legal assistant workflows, including general legal questions and answers, retrieval-augmented generation, multi-step reasoning, and scenario-based problem solving tailored to Vietnamese law. By providing a standardized, transparent, and cognitively informed evaluation framework, VLegal-Bench establishes a solid foundation for assessing LLM performance in Vietnamese legal contexts and supports the development of more reliable, interpretable, and ethically aligned AI-assisted legal systems. To facilitate access and reproducibility, we provide a public landing page for this benchmark at https://vilegalbench.cmcai.vn/.

Vision-Language Foundation Models (VLMs), trained on large-scale multimodal datasets, have driven significant advances in Artificial Intelligence (AI) by enabling rich cross-modal reasoning. Despite their success in general domains, applying these models to medical imaging remains challenging due to the limited availability of diverse imaging modalities and multilingual clinical data. Most existing medical VLMs are trained on a subset of imaging modalities and focus primarily on high-resource languages, thus limiting their generalizability and clinical utility. To address these limitations, we introduce a novel Vietnamese-language multimodal medical dataset consisting of 2,757 whole-body PET/CT volumes from independent patients and their corresponding full-length clinical reports. This dataset is designed to fill two pressing gaps in medical AI development: (1) the lack of PET/CT imaging data in existing VLMs training corpora, which hinders the development of models capable of handling functional imaging tasks; and (2) the underrepresentation of low-resource languages, particularly the Vietnamese language, in medical vision-language research. To the best of our knowledge, this is the first dataset to provide comprehensive PET/CT-report pairs in Vietnamese. We further introduce a training framework to enhance VLMs' learning, including data augmentation and expert-validated test sets. We conduct comprehensive experiments benchmarking state-of-the-art VLMs on downstream tasks. The experimental results show that incorporating our dataset significantly improves the performance of existing VLMs. We believe this dataset and benchmark will serve as a pivotal step in advancing the development of more robust VLMs for medical imaging, especially for low-resource languages and clinical use in Vietnamese healthcare. The source code is available at https://github.com/AIoT-Lab-BKAI/ViPET-ReportGen.

Pathology vision-language models are advancing rapidly, yet existing benchmarks remain focused on human tissue, particularly oncology, leaving non-human pathology largely unaddressed. This gap is especially important in toxicologic pathology, where microscopic tissue examination of laboratory animals is a core component of preclinical drug safety assessment. To address it, we introduce VIPER, the first expert-curated benchmark for vision-language model evaluation in toxicologic pathology. VIPER contains 1,251 questions associated with 419 H&E-stained rat histology images across seven organ systems, covering multiple-choice, KPrim, and free-text formats. All questions were curated and validated by board-certified veterinary pathologists. In total, we benchmarked 16 models, including two newly introduced veterinary-pathology models, seven human pathology-specialized models, and seven general-purpose frontier models. The results identify a substantial domain gap between veterinary and human pathology, expose the risk of over-diagnosis of normal tissue in frontier models, and show that domain-specific training remains critical for visually grounded predictions. VIPER data and evaluation code are available at https://github.com/mahmoodlab/viper.

Test-time scaling can improve large language model reasoning by generating and combining multiple candidate responses. In sampling-based methods, the inference budget is often described by the number of generated candidates, N. However, N tells us how many candidates are generated, not how they are executed. The same candidate budget can be produced in one batched generation call or split across several sequential calls with smaller batch sizes. We first study the effect of increasing N on reasoning accuracy using Phi-3-mini and Qwen2.5-1.5B on 500 GSM8K prompts. As expected, increasing N from 1 to 8 improves accuracy by 8.4 percentage points for Phi-3-mini and 18.4 points for Qwen2.5-1.5B. However, accuracy alone does not show the systems cost of using a larger candidate budget. We therefore fix N = 8 and compare four generation schedules: 1x8, 2x4, 4x2, and 8x1, where axb denotes a generation calls with b candidates per call. We measure latency, throughput, GPU-hours, and gross GPU-device energy while keeping the total candidate count fixed. On A100 GPUs, eight serial calls use 4.64-4.86x as much gross GPU-device energy and have 5.77-6.12x the P95 latency of one batched call with eight candidates. The same pattern appears across three independently scheduled A100 nodes per model and in short-output SciQ/V100 experiments. These results show that candidate count alone is not enough to describe the systems cost of multi-candidate test-time scaling. When candidates are independent and memory allows it, fewer generation calls with larger batch sizes are more efficient. Evaluations should therefore report not only candidate count and accuracy, but also generation schedule and GPU-level systems metrics.

Medication reconciliation at clinical handoffs is a high-stakes, error-prone process. Large language models are increasingly proposed to assist with this task using FHIR-structured patient records, but a fundamental and largely unstudied variable is how the FHIR data is serialised before being passed to the model. We present the first systematic comparison of four FHIR serialisation strategies (Raw JSON, Markdown Table, Clinical Narrative, and Chronological Timeline) across five open-weight models (Phi-3.5-mini, Mistral-7B, BioMistral-7B, Llama-3.1-8B, Llama-3.3-70B) on a controlled benchmark of 200 synthetic patients, totalling 4,000 inference runs. We find that serialisation strategy has a large, statistically significant effect on performance for models up to 8B parameters: Clinical Narrative outperforms Raw JSON by up to 19 F1 points for Mistral-7B (r = 0.617, p < 10^{-10}). This advantage reverses at 70B, where Raw JSON achieves the best mean F1 of 0.9956. In all 20 model and strategy combinations, mean precision exceeds mean recall: omission is the dominant failure mode, with models more often missing an active medication than fabricating one, which changes how clinical safety auditing priorities should be set. Smaller models plateau at roughly 7-10 concurrent active medications, leaving polypharmacy patients, the patients most at risk from reconciliation errors, systematically underserved. BioMistral-7B, a domain-pretrained model without instruction tuning, produces zero usable output in all conditions, showing that domain pretraining alone is not sufficient for structured extraction. These results offer practical, evidence-based format recommendations for clinical LLM deployment: Clinical Narrative for models up to 8B, Raw JSON for 70B and above. The complete pipeline is reproducible on open-source tools running on an AWS g6e.xlarge instance (NVIDIA L40S, 48 GB VRAM).

We reassess the realistic discovery reach of Solar-System experiments for dark energy (DE) and dark matter (DM), making explicit the bridge from cosmology-level linear responses to local, screened residuals. In scalar-tensor frameworks with a universal conformal coupling A(phi) and chameleon/Vainshtein screening, we map cosmological responses {mu(z,k),Sigma(z,k)} inferred by DESI and Euclid to thin-shell or Vainshtein residuals in deep Solar potentials Phi_N. We emphasize a two-branch strategy. In a detection-first branch, a verified local anomaly -- an Einstein equivalence principle (EEP) violation, a Shapiro-delay signal with |gamma-1|simfewtimes 10^{-6}, an AU-scale Yukawa tail, or a ultralight DM (ULDM) line in clocks/atom interferometers in space (AIS) -- triggers a joint refit of cosmology and Solar-System data under a common microphysical parameterization {V(phi),A(phi)}. In a guardrail branch, Solar-System tests enforce constraints (EEP; PPN parameters gamma,beta; and dot G/G) and close unscreened or weakly screened corners indicated by cosmology. We forecast, per conjunction, |gamma-1|lesssim (2-5)times 10^{-6} (Ka-/X-band or optical Shapiro), eta_{EEP}sim (1--10)times 10^{-17} (drag-free AIS), |dot G/G|sim(3-5)times10^{-15},yr^{-1} (sub-mm-class LLR), a uniform ~2x tightening of AU-scale Yukawa/DM-density bounds, and (3-10)times improved ULDM-coupling reach from clocks. For a conformal benchmark, mu_{ lin,0}=0.10 implies chisimeq mu_{lin,0/2} and a Sun thin shell Delta R/Rlesssim (1/3chi)|gamma-1|/2=2.4times 10^{-3} at |gamma-1|=5times 10^{-6}; Vainshtein screening at 1 AU yields |gamma-1|lesssim 10^{-11}, naturally below near-term reach. We recommend a cost-effective guardrail+discovery portfolio with explicit triggers for escalation to dedicated missions.

Transformers have achieved great success in numerous NLP tasks but continue to exhibit notable gaps in multi-step factual reasoning, especially when real-world knowledge is sparse. Recent advances in grokking have demonstrated that neural networks can transition from memorizing to perfectly generalizing once they detect underlying logical patterns - yet these studies have primarily used small, synthetic tasks. In this paper, for the first time, we extend grokking to real-world factual data and address the challenge of dataset sparsity by augmenting existing knowledge graphs with carefully designed synthetic data to raise the ratio phi_r of inferred facts to atomic facts above the threshold required for grokking. Surprisingly, we find that even factually incorrect synthetic data can strengthen emergent reasoning circuits rather than degrade accuracy, as it forces the model to rely on relational structure rather than memorization. When evaluated on multi-hop reasoning benchmarks, our approach achieves up to 95-100% accuracy on 2WikiMultiHopQA - substantially improving over strong baselines and matching or exceeding current state-of-the-art results. We further provide an in-depth analysis of how increasing phi_r drives the formation of generalizing circuits inside Transformers. Our findings suggest that grokking-based data augmentation can unlock implicit multi-hop reasoning capabilities, opening the door to more robust and interpretable factual reasoning in large-scale language models.

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.

This paper assesses the ability of large language models (LLMs) to translate texts that include inter-sentential dependencies. We use the English-French DiscEvalMT benchmark (Bawden et al., 2018) with pairs of sentences containing translation challenges for pronominal anaphora and lexical cohesion. We evaluate 12 LLMs from the DeepSeek-R1, GPT, Llama, Mistral and Phi families on two tasks: (1) distinguish a correct translation from a wrong but plausible one; and (2) generate a correct translation. We compare prompts that encourage chain-of-thought reasoning with those that do not. The best models take advantage of reasoning and reach about 90% accuracy on the first task and COMET scores of about 92% on the second task, with GPT-4, GPT-4o and Phi standing out. Moreover, we observe a "wise get wiser" effect: the improvements through reasoning are larger for models that already perform well without reasoning.

Today's reasoning models use thinking tokens to attain stronger performance on benchmarks than their instruction-tuned counterparts. It is also generally believed that this more "deliberative" mode should improve alignment and safety, by providing the model a safe space to consider whether its planned answer to a request violates its safety principles. We present evidence that this intuition is not always correct. Across frontier open-weight reasoning models spanning GPT-OSS, Qwen, Olmo, and Phi families, we find that the eventual refusal/compliance outcome is already strongly predictable via a trained head on the first token's hidden representation (0.84-0.95 AUROC and sim88% balanced accuracy for predicting refusal/compliance) before any visible thinking. The thinking process turns out to be more akin to prefix completion than to deliberative revision, with the final outcome rarely changing after the first sim20% of thinking, despite giving the appearance of deliberation at the text level (sim74% of text-level deliberations occur when the response distribution is already locked to one refusal/compliance side). We also find that existing inference-time and training-based safety interventions, despite being motivated by the goal of inducing deliberation, largely shift model behavior toward over-refusal while suppressing already-scarce deliberation signals. Our results suggest that safety behavior in current reasoning models is much less deliberative than commonly assumed, and highlight the need for methods that induce real safety deliberation.