arXiv AI

arXiv:2606.25034v3 Announce Type: replace-cross Abstract: General-purpose models often struggle to reliably identify and understand real-world multimodal risks, largely due to the inherent multimodal adversarial nature of content and AI safety. We present Yuvion VL, a family of multimodal large language models purpose-built for content and AI safety, with both instruction-tuned and reasoning-oriented variants. Yuvion VL addresses this gap by treating safety as an inherently adversarial and multimodal problem and designing the entire pipeline around adversarial robustness. For data construction, we develop an automated pipeline integrating adversarial-aware data synthesis with multi-stage quality control, producing large-scale, high-quality multimodal samples augmented with domain knowledge and reasoning annotations. For training, we adopt a three-stage pipeline that includes continued pretraining for risk-concept cross-modal alignment, instruct post-training for production-grade safety tasks, and reasoning post-training for enhanced interpretability and performance in complex tasks. We further introduce Confuse-then-Contrast Fine-Tuning, a contrastive framework that mines model-specific confusions and constructs multi-image contrastive groups to enforce explicit discrimination of fine-grained visual-semantic elements, enabling the model to distinguish between visually similar cases with different safety implications in adversarial safety tasks. To support rigorous evaluation, we further introduce Yuvion VL RiskEval (YVRE), a collection of benchmarks covering diverse open and internal evaluations, with a focus on content and AI safety, adversarial robustness, and real-world capability requirements. Experiments show that Yuvion VL-32B achieves industry-leading safety performance, surpassing comparably sized open-source models and best closed-source commercial models, while maintaining comparable general capabilities.

arXiv:2606.22113v2 Announce Type: replace-cross Abstract: Generalizing manipulation policies across robot embodiments remains difficult because standard policies tie the interactions needed for a task to the actions of a particular robot. We study zero-shot cross-embodiment manipulation, where a policy trained on source embodiments must be deployed on a structurally different target embodiment without additional task demonstrations. We introduce Kinematic Interaction Transfer across Embodiments (KITE), which decouples manipulation into embodiment-agnostic interaction prediction and embodiment-specific kinematic decoding, connected through a learned representation of interaction intent based on contact patterns. An interaction policy is trained to predict the intent using demonstration data only collected on a source robot, and shared across morphologically different target robots. A kinematic decoder realizes these intents by learning from synthetic data generated from each embodiment's kinematic model. Through such decoupling, KITE can robustly transfer learned skills to a new embodiment given only its kinematic model, without recollecting any task-specific demonstration data. We evaluate KITE on four manipulation tasks in simulation and the real world spanning transfer between parallel grippers, dexterous hands, and composite embodiments. KITE successfully achieves zero-shot transfer to structurally different target embodiments, and consistently outperforms state-of-the-art baselines. Project website: https://kite-manip.github.io/.

arXiv:2606.15555v3 Announce Type: replace-cross Abstract: In large language model (LLM) serving, each request accumulates persistent graphics processing unit (GPU) memory during service as its key-value cache grows with every generated token. Under high concurrency, aggregate memory usage therefore increases endogenously over time: the service process itself creates future capacity pressure. When memory capacity is exceeded, systems evict active requests, discarding cached state and restarting them later, which wastes computation and reduces throughput. We develop a discrete-time dynamical model of memory-constrained LLM inference that captures admission, memory growth, and eviction under continuous batching. In the saturated-input regime, the system admits both eviction-free fixed points and limit cycles with evictions. For homogeneous workloads, we show that the eviction-free equilibrium is unstable and that, except for a Lebesgue-measure-zero exact-capture set, the system converges to a unique worst-case limit cycle that is asymptotically stable outside this exceptional set, with throughput losses as large as 50%. For heterogeneous workloads, we prove a stability criterion in the two-class common-input setting and explain how the survival-polynomial mechanism generalizes to multiple classes and heterogeneous-input lengths. Under an input-dominated scaling regime, coprime decoding lengths stabilize the eviction-free equilibrium, while non-coprime lengths create synchronized modes that drive instability. These results characterize when workload heterogeneity desynchronizes completions and helps stabilize memory-constrained serving. More broadly, we identify service-induced congestion as a structural instability mechanism and derive scheduling design principles for sustaining high throughput.

arXiv:2606.10180v2 Announce Type: replace-cross Abstract: We introduce flow control of vision-language-action (VLA) models, a simple and effective way to steer VLA actions in real time through generic inputs, such as a keyboard. This method can be used out-of-the-box and does not require retraining or fine-tuning VLAs. It enables relatively crude user inputs to steer a VLA to align with user intent. Flow control writes the user input into the initial condition of the VLA's flow matching action expert. Because the flow is deterministic, integration retains information about the user input while moving it toward the action distribution learned during training, so that the generated actions reflect the user's intent (high fidelity) and stay close to the VLA's own unsteered actions (high conformity). We demonstrate that flow control has many desirable properties: (1) flow control accurately and responsively steers robot actions with user inputs, (2) it is robust to suboptimal user inputs, (3) it enables users to steer VLAs to achieve significantly higher success rates and faster task completion, and (4) fine-tuning a VLA on flow control trajectories improves the autonomous policy. Together, these results provide a simple and intuitive way for users to help steer VLA actions, increasing task performance. Project page and videos: https://jcykao.github.io/flowcontrol/.

arXiv:2605.30251v2 Announce Type: replace-cross Abstract: Large language models (LLMs) often solve a task when all instructions are given in a single prompt, but fail when the same information is revealed gradually across turns. When a clean FULL prompt and a RAW-SHARDED conversation contain the same complete user evidence, the model should still arrive at the same answer. We argue that a key reason for this gap is self-anchored drift: responses produced under partial information introduce unsupported assumptions, and those assumptions later distort the final answer. To reduce this effect, we propose Canonical-Context On-Policy Distillation (CCOPD). During training, the same base model is used in two roles: a frozen teacher conditioned on the clean FULL prompt and a trainable student that receives the same evidence incrementally through a multi-turn conversation; CCOPD aligns the student's behavior on its own trajectories with the teacher's canonical full-context behavior. Trained only on math problem conversations, CCOPD yields a 32\% average relative improvement in RAW-SHARDED performance over the original base model across math and five zero-shot out-of-domain task families, while largely preserving full-context performance. Further analyses suggest that CCOPD strengthens grounding in user evidence and reduces sensitivity to contamination from earlier assistant turns.

arXiv:2605.29458v2 Announce Type: replace-cross Abstract: Large language models are increasingly asked to stand in for particular people, and interviewing the person is the natural way to tell a model who they are. We test what a one-hour LLM-led adaptive interview tells a model about an individual's value-laden decisions. 46 adults completed such an interview and then answered 25 dilemma questions, which are the ground truth. We predict their answers from different parts of the transcript, compare adaptive follow-up questions with fixed ones on the same participants, and benchmark every variant against a model given no information about the person and against the majority answer. No part of the interview is reliably better than either baseline, and adaptive follow-ups are no better than fixed ones. Two audits of the transcripts show why. For 85% of the errors an auditor finds no evidence that distinguishes the person's choice from the alternatives, and where a transcript does contain related evidence, the option it favours matches the person's own answer less often than the majority answer does. The same transcripts do improve recovery of the person's MBTI type. Interviews of this kind elicit traits, not decision policies. We discuss what decision-aware elicitation would require and why persona agents should not be delegated value-laden decisions, and we release the consenting participants' transcripts with all predictions.

arXiv:2605.21187v2 Announce Type: replace-cross Abstract: As distributed model training scales to span hundreds of thousands of GPUs, scale-out networks face unprecedented performance and efficiency demands. NVIDIA Spectrum-X Ethernet has been designed from the ground up to achieve predictable and stable network performance with high utilization and low latency. This paper presents the Spectrum-X multiplane architecture, which replaces hierarchical depth with topological parallelism, and introduces hardware-accelerated load balancing in NICs and switches as the key architectural approach to provide fast reaction to highly dynamic network conditions at the microsecond timescales that AI training workloads demand. We describe the motivation, design principles, evaluation methodology and performance on state-of-the-art benchmarks, as well as the lessons we learned from deploying and debugging Spectrum-X networks in large-scale systems. Our evaluation highlights production-grade AI infrastructure performance across four core dimensions: 98% of the theoretical line rate with low jitter-free latency; strong cross-tenant isolation for concurrent workloads; robust, capacity-proportional bisection bandwidth and 7% latency increase for 10% fabric link failures; and rapid reaction to host and fabric link flaps during LLM training workloads.

arXiv:2605.20708v4 Announce Type: replace-cross Abstract: Diffusion Transformers (DiTs) have become a de facto backbone of modern visual generation, and nearly every major axis of their design -- tokenization, attention, conditioning, objectives, and latent autoencoders -- has been extensively revisited. The residual stream that governs how information accumulates across layers, however, has been directly inherited from the original Transformer. In this paper, we present a systematic empirical analysis of cross-layer information flow in DiTs, jointly along depth and denoising timestep, and identify three concrete symptoms of traditional residual addition, namely monotonic forward magnitude inflation, sharp backward gradient decay, and pronounced block-wise redundancy. Motivated by this diagnosis, we propose Diffusion-Adaptive Routing (DAR), a drop-in residual replacement that performs learnable, timestep-adaptive, and non-incremental aggregation over the history of sublayer outputs. Moreover, the proposed DAR is compatible with many modern Transformer enhancement methods, such as REPA. On ImageNet $256\times256$, DAR improves SiT-XL/2 by $2.11$ FID ($7.56$ vs. $9.67$) and matches the baseline's converged quality with $8.75\times$ fewer training iterations. Stacked on top of REPA, it yields a $2\times$ training acceleration in the early stage, suggesting cross-layer information routing as an underexplored design axis in diffusion modeling, one that operates orthogonally to existing representation-alignment objectives. Beyond pretraining, DAR can also be applied during the fine-tuning stage of large-scale T2I models and preserves high-frequency details during Distribution Matching Distillation. Code is available at https://github.com/cg233666/attention-residuals-on-dit.

arXiv:2605.19294v3 Announce Type: replace-cross Abstract: Vision-Language-Action (VLA) policies increasingly rely on asynchronous inference to hide model latency behind ongoing robot motion. During inference, the robot continues executing its current action chunk while the next chunk is generated from the observation available at inference start. By the time the new chunk is executed, the robot or scene may have changed, so an action that fits the prediction-time state can be misaligned with the execution-time state. Training-time methods modify the policy input or supervised target based on the expected execution delay, but do not directly supervise how actions should change as the scene evolves between prediction and execution. Runtime repair methods improve chunk continuity or modify actions during deployment, but can still leave actions aligned with the prediction-time state rather than the execution-time state. Our key insight is that execution-time observations are unavailable when the online prediction is made but are available in recorded trajectories during offline training, making the prediction-execution mismatch observable during training. We propose DEFLECT, an offline post-training framework that turns this temporal contrast into preference supervision for a policy acting from stale inputs. Across Kinetix, LIBERO, and three real-robot tasks, DEFLECT improves delay robustness across the evaluated VLA policies and settings. At the longest supported delay, it improves success by 4.2 percentage points over VLASH on Kinetix and by 22.7 points over naive asynchronous execution with GR00T N1.7 on LIBERO.

arXiv:2605.18904v2 Announce Type: replace-cross Abstract: Model merging consolidates independently fine-tuned checkpoints without joint training or access to task data. Dynamic methods retain task-specific capacity, but must divide a limited storage budget between shared and expert parameters. We propose DiDi-Merging, which jointly allocates low-rank capacity to these components through differentiable rank selection, followed by data-free parameter-space refinement. On the eight-task ViT-B/32 benchmark, the small configuration reaches 89.3% average accuracy (98.7% normalized retention) at 1.24x the storage of one backbone, including that backbone; the medium configuration reaches 89.8% at 1.4x. Removing refinement reduces the medium configuration by 4.3 percentage points, while static-merge plus uniform-recovery controls are 4.7-8.9 points below the small configuration. Additional evaluations cover language and multimodal models at their separately reported budgets. The main results assume known task identity; practical routing and accuracy under larger task counts at fixed capacity remain distinct evaluation questions. Code is available at https://github.com/WanyuGroup/NeurIPS2026-DiDi-Merging.

arXiv:2605.17826v2 Announce Type: replace-cross Abstract: Do Vision-Language Models (VLMs) count what they see, or what they know about the world? We introduce CounterCount, a diagnostic dataset of paired factual and counterfactual (CF) images, in which the CF image contains an atypical count of a subject's attribute (e.g., a rabbit with four ears), together with annotations localizing the counted attributes. Humans answer both types of images nearly perfectly ($98.9\%$ and $97.5\%$), but across 16 VLMs (14 open-weight and 2 closed-source), average accuracy drops from $91.6\%$ on factual to $49.5\%$ on CF images, with most errors stemming from responses that match the canonical count, therefore revealing a bias towards knowledge priors rather than reliance on visual information. Based on these findings, we perform an analysis that reveals two interesting insights behind this gap. First, models exhibit a ``System-1''-like response behavior, falling back on canonical counts after a shallow recognition of the subject. Second, model activations partially contain the visual evidence necessary to answer correctly, yet models underutilize it, attending less to the counted attributes when they provide biased answers. Input manipulations that break visual context, introduce cueing or reasoning, or direct amplification of the model's internal attention to relevant regions can reduce such biases. CounterCount goes beyond showing that VLMs fall back on knowledge priors, revealing that once a familiar context is recognized, models favor prior knowledge over available visual evidence rather than failing to encode that evidence.

arXiv:2605.10765v2 Announce Type: replace-cross Abstract: Multimodal Large Language Models (MLLMs) achieve strong performance through instruction tuning, yet real-world deployment often requires continual capability expansion across sequential tasks. In such scenarios, Multimodal Continual Instruction Tuning (MCIT) aims to acquire new capabilities while limiting catastrophic forgetting. Existing methods mainly follow a module-composition paradigm: they maintain task-level prompts or LoRA experts and dynamically route or aggregate a subset of them at inference. However, samples within the same task can still differ substantially in visual scenes, question intents, and reasoning demands. This motivates instance-level adaptation to individual query-image pairs rather than only selecting or combining task-level modules. To this end, we propose DRAPE (Dynamic Cross-Modal Prompt Generation), a prompt-learning framework that synthesizes continuous instance-specific soft prompts for MCIT. Instead of selecting prompts from a fixed pool, DRAPE derives prompt queries from the textual instruction and cross-attends to visual patch features, producing prompts conditioned on query-image pairs that are prepended to the frozen LLM. To mitigate forgetting during sequential updates, DRAPE applies null-space gradient projection to the shared projector and uses CLIP-based prototype routing for task-label-free generator selection at inference. Extensive experiments on MCIT benchmarks show that DRAPE achieves state-of-the-art performance among representative prompt-based and LoRA-based continual-learning baselines. Code is available at: https://github.com/LAMDA-CL/NeurIPS2026-Drape.

arXiv:2605.10641v2 Announce Type: replace-cross Abstract: Large Vision-Language Models (VLMs) are successful in addressing a multitude of vision-language understanding tasks, such as Visual Question Answering (VQA), but their memory and compute requirements remain a concern for practical deployment. A promising class of techniques for mitigating this concern is Knowledge Distillation, where knowledge from a high-capacity Teacher network is transferred to a considerably smaller Student network. However, the capacity gap between the two networks is both a blessing and a curse: the smaller the Student network, the better its efficiency, and the stronger (typically, also larger) the Teacher, the more knowledge it carries; yet, beyond a point, the larger capacity gap between the two leads to worse knowledge transfer. To counter this effect, we propose a bottom-up cascaded knowledge distillation (CKD) framework. Instead of treating knowledge transfer as an activity involving one high-capacity Teacher (or an ensemble of such), inspired by human formal education systems, we introduce one (potentially, more) additional Teacher(s) of intermediate capacity that gradually bring the Student network to the next level, where the next (higher-capacity) Teacher can take over. We provide a theoretical analysis to motivate why progressive reduction of the Teacher-Student gap may improve distillation. We apply the proposed framework on models built upon the LLaVA methodology and evaluate the derived models on seven standard, publicly available VQA benchmarks, demonstrating competitive performance with recent compact VLMs. Source code for cascaded distillation and trained models are available at https://github.com/IDT-ITI/LLaVA-CKD.

arXiv:2603.26246v2 Announce Type: replace-cross Abstract: Standard LLM-based speech recognition systems process utterances in isolation, limiting their ability to leverage conversational context. Here we study whether multimodal context from prior turns improves LLM-based ASR and how to represent that context efficiently. We find that, after supervised multi-turn training, conversational context mainly helps with the recognition of contextual entities. However, conditioning on raw context is expensive because the prior-turn audio embedding sequence grows rapidly with conversation length. To address this, we propose Abstract Compression, a learned latent abstraction of prior-turn audio that replaces each prior-turn audio sequence with a fixed number of latent tokens while retaining corresponding transcripts explicitly. On both in-domain and out-of-domain test sets, the compression-enabled model recovers part of the gains of raw-context conditioning with a smaller prior-turn audio footprint. We also provide targeted analyses of the compression setup and its tradeoffs.

arXiv:2603.25495v2 Announce Type: replace-cross Abstract: Accurate short-term air-quality forecasting is essential for public health protection and urban management, yet many recent forecasting frameworks rely on complex, data-intensive, and computationally demanding models. This study investigates whether lightweight and interpretable forecasting approaches can provide competitive performance for hourly PM2.5 prediction in Beijing, China. Using multi-year pollutant and meteorological time-series data, we developed a leakage-aware forecasting workflow that combined chronological data partitioning, preprocessing, feature selection, and exogenous-driver modeling under the Perfect Prognosis setting. Three forecasting families were evaluated: SARIMAX, Facebook Prophet, and NeuralProphet. To assess practical deployment behavior, the models were tested under two adaptive regimes: weekly walk-forward refitting and frozen forecasting with online residual correction. Results showed clear differences in both predictive accuracy and computational efficiency. Under walk-forward refitting, Facebook Prophet achieved the strongest completed performance, with an MAE of $37.61$ and an RMSE of $50.10$, while also requiring substantially less execution time than NeuralProphet. In the frozen-model regime, online residual correction improved Facebook Prophet and SARIMAX, with corrected SARIMAX yielding the lowest overall error (MAE $32.50$; RMSE $46.85$). NeuralProphet remained less accurate and less stable across both regimes, and residual correction did not improve its forecasts. Notably, corrected Facebook Prophet reached nearly the same error as its walk-forward counterpart while reducing runtime from $15$ min $21.91$ sec to $46.60$ sec. These findings show that lightweight additive forecasting strategies can remain highly competitive for urban air-quality prediction, offering a practical balance between accuracy, interpretability, ...

arXiv:2603.18280v4 Announce Type: replace-cross Abstract: Current alignment evaluation mostly measures whether models encode dangerous concepts and whether they refuse harmful requests. Both miss the layer where alignment often operates: routing from concept detection to behavioral policy. We study political censorship in Chinese-origin language models as a natural experiment, using probes, surgical ablations, and behavioral tests across nine open-weight models from five labs. Three findings follow. First, probe accuracy alone is non-diagnostic: political probes, null controls, and permutation baselines can all reach 100%, so held-out category generalization is the informative test. Second, surgical ablation reveals lab-specific routing. Removing the political-sensitivity direction eliminates censorship and restores accurate factual output in most models tested, while one model confabulates because its architecture entangles factual knowledge with the censorship mechanism. Cross-model transfer fails, indicating that routing geometry is model- and lab-specific. Third, refusal is no longer the dominant censorship mechanism. Within one model family, hard refusal falls to zero while narrative steering rises to the maximum, making censorship invisible to refusal-only benchmarks. These results support a three-stage descriptive framework: detect, route, generate. Models often retain the relevant knowledge; alignment changes how that knowledge is expressed. Evaluations that audit only detection or refusal therefore miss the routing mechanism that most directly determines behavior.

arXiv:2603.17651v2 Announce Type: replace-cross Abstract: Generative inbetweening (GI) seeks to synthesize realistic intermediate frames between the first and last keyframes beyond mere interpolation. As sequences become sparser and motions larger, previous GI models struggle with inconsistent frames with unstable pacing and semantic misalignment. Since GI involves fixed endpoints and numerous plausible paths, this task requires additional guidance gained from the keyframes and text to specify the intended path. Thus, we give semantic and temporal guidance from the keyframes and text onto each intermediate frame through Keyframe-anchored Attention Bias. We also better enforce frame consistency with Rescaled Temporal RoPE, which allows self-attention to attend to keyframes more faithfully. TGI-Bench, the first benchmark specifically designed for text-conditioned GI evaluation, enables challenge-targeted evaluation to analyze GI models. Without additional training, our method achieves state-of-the-art frame consistency, semantic fidelity, and pace stability for both short and long sequences across diverse challenges.

arXiv:2603.17372v3 Announce Type: replace-cross Abstract: Large vision-language models (VLMs) often exhibit weakened safety alignment with the integration of the visual modality. Even when text prompts contain explicit harmful intent, adding an image can substantially increase jailbreak success rates. In this paper, we observe that VLMs can clearly distinguish benign inputs from harmful ones in their representation space. Moreover, even among harmful inputs, jailbreak samples form a distinct internal state that is separable from refusal samples. These observations suggest that jailbreaks do not arise from a failure to recognize harmful intent. Instead, the visual modality shifts representations toward a specific jailbreak state, thereby leading to a failure to trigger refusal. To quantify this transition, we identify a jailbreak direction and define the jailbreak-related shift as the component of the image-induced representation shift along this direction. Our analysis shows that the jailbreak-related shift reliably characterizes jailbreak behavior, providing a unified explanation for diverse jailbreak scenarios. Finally, we propose a defense method that enhances VLM safety by removing the jailbreak-related shift (JRS-Rem) at inference time. Experiments show that JRS-Rem provides strong defense across multiple scenarios while preserving performance on benign tasks. Code is available at https://github.com/LeeQueue513/JRS-Rem.

arXiv:2603.11413v5 Announce Type: replace-cross Abstract: A recent Nature Medicine study reported that ChatGPT Health under-triages 51.6% of emergencies and concluded that consumer-facing AI triage poses safety risks. Its protocol, however, was an exam-style scaffold (forced A/B/C/D output, knowledge suppression, no clarifying questions) unlike how consumers use health chatbots. We ask how much of the headline error rate is a property of the measurement rather than of the models. In a first, mechanistic study, five frontier LLMs on a 17-scenario bank scored 6.4 points higher under naturalistic patient-style messages than under the constrained scaffold (p=0.015). In a second, faithful replication we ran the authors' own 60 released vignettes through six frontier models under four matched formats, with clinician validation of the rewrites and a blinded clinician audit of the LLM adjudicators. Here the direction reversed: free-text rewrites scored slightly below the exact prompt (78.7% vs 81.8%, p=0.020), removing only the answer scaffold changed little (80.6% vs 81.4%, p=0.81), and naturalistic input with a forced categorical answer beat both the exact prompt (84.4%, p=0.023) and free text (p=0.0001). On the four vignettes defining the original emergency rate, under-triage was 17% with the exact scaffold, 50% in free text, and 21% with a forced letter on the same message; every free-text "under-triage" was a same-day recommendation scored C rather than D, and 60% made escalation conditional on information the patient was asked to check, behavior a single-turn benchmark cannot score. The headline rate did not reproduce as a stable cross-model finding; much of it depends on output format and on mapping prose onto a four-point scale. Each format elicits a separate response, so the design measures format sensitivity, not format versus capability. Safety claims should report sensitivity to input wording, output format and adjudication.

arXiv:2603.00889v2 Announce Type: replace-cross Abstract: Large Language Models (LLMs) have exhibited remarkable reasoning capabilities, largely enabled by supervised fine-tuning (SFT)- and reinforcement learning (RL)-based post-training on high-quality reasoning data. However, reproducing and extending these capabilities in open and scalable settings is hindered by three data-centric challenges: (1) the cold-start problem, arising from the lack of seed datasets with detailed, long Chain-of-Thought (CoT) trajectories needed to initialize reasoning policies; (2) limited domain coverage, as most existing open-source reasoning datasets are concentrated in mathematics, with limited coverage of broader scientific disciplines; and (3) the annotation bottleneck, where the difficulty of frontier-level reasoning tasks makes reliable human annotation prohibitively expensive or infeasible. To address these challenges, we introduce CHIMERA, a compact synthetic reasoning dataset of 9K samples designed to support generalizable reasoning across domains. CHIMERA is constructed with three key properties: (1) it provides rich, long CoT reasoning trajectories synthesized by state-of-the-art reasoning models; (2) it has broad and structured coverage, spanning 8 major scientific disciplines and over 1K fine-grained topics organized via a model-generated hierarchical taxonomy; and (3) it employs a fully automated, scalable evaluation pipeline that uses strong reasoning models to cross-validate both problem validity and answer correctness, removing the reliance on human annotations. We use CHIMERA to post-train a 4B Qwen3 model with SFT and RL. Despite the dataset's modest size, the resulting model achieves strong performance on a suite of challenging reasoning benchmarks, including GPQA-Diamond, AIME 24/25/26, HMMT 25 and Humanity's Last Exam, approaching or matching the performance of substantially larger models such as DeepSeek-R1 and Qwen3-235B.