Test-time scaling (TTS) reliably improves reasoning in large language models, but whether it transfers to small open vision-language models remains unclear. We examine this on EXAMS-V, a multilingual visual multiple-choice benchmark, comparing self-consistency, describe-then-reason with PRM-guided beam search, and two post-hoc selectors across Qwen2.5-VL-7B-Instruct and Qwen3.5-4B. What matters is the conditions under which TTS runs, not the search or verification machinery. The largest factor is parseability: an early prompt format left many chains reasoning correctly yet never committing to an answer letter, which a standard answer cue and a guided repair step largely remove. A larger decoding budget removes the rest: raising the per-chain token limit from 1k to 2k recovers 3.7 pp, whereas sampling more chains (8 to 16) adds only 0.15 pp. Once chains have room to finish, elaborate methods contribute little: PRM-guided beam search trails plain self-consistency by 0.39 pp at over eight times the cost, and neither a training-free generative critic nor a trained multimodal PRM beats majority vote across both policies. The largest gain comes instead from the policy model itself (+11.4 pp). Our best configuration reaches 84.1% on the held-out ImageCLEF 2026 test split, ranking first on the Visual MCQ leaderboard.
Large-scale autonomous-driving datasets contain vast numbers of recorded scenarios, creating a need for efficient retrieval methods that can identify situations similar to a given query. Existing approaches typically rely on either visual representations or motion-based descriptions, making it difficult to understand their relative strengths and limitations for scenario retrieval. In this work, we present a multimodal framework for autonomous-driving scenario retrieval that combines visual and trajectory-based representations within a unified retrieval pipeline. We investigate two trajectory-based approaches: Exo-Trajectory, an explicit matching method based on surrounding-agent motion, and ScenarioFormer, a transformer-based representation learned from object trajectories using contrastive learning. We compare these approaches against strong vision-based baselines and analyze their behavior across a diverse set of driving scenarios. Experimental results show that trajectory representations provide strong retrieval performance for motion-centric events such as cut-ins, turning maneuvers, and traffic queueing, while visual embeddings excel when appearance cues are informative. Most importantly, combining visual and trajectory information consistently improves retrieval quality, yielding the best overall performance. These findings demonstrate that appearance and motion capture are complementary notions of scenario similarity and motivate multimodal retrieval systems for autonomous-driving data mining, dataset curation, and scenario-based validation.
We present Soofi S 30B-A3B, a sovereign, open-source Mixture-of-Experts (MoE) hybrid Mamba Transformer foundation model for German and English. Its hybrid design activates only 3B of 30B parameters per token and keeps the inference cache near-constant as context grows, giving it a decisive throughput advantage over dense models for long-context, high-concurrency deployment. Pretrained on roughly 27 trillion tokens with deliberately up-weighted German, Soofi S matches dense 14 to 27B models on aggregate English and German benchmarks while achieving the best code aggregates in both languages among 17 open base models, and outperforms every European sovereign baseline in our comparison, including ones far larger in active parameters. Among fully open models, Soofi S obtains the highest English and German evaluation scores, ahead of Olmo 3 32B and Apertus 70B. Soofi S was built end-to-end on the German Industrial AI Cloud, a sovereign HPC scale AI infrastructure operated by Deutsche Telekom in Munich. Soofi S will be released under highly permissive, open-access terms: weights, selected intermediate checkpoints, full per-source data accounting, hyperparameters, and training and evaluation code. Where source licenses permit, data-construction artifacts are released under permissive licenses; commercially licensed sources are documented with aggregate statistics and exact mixture accounting.
Ambivalence and hesitancy are subtle behavioral states that are expressed through a combination of verbal content, facial behavior, visual context, and acoustic cues. Effective recognition therefore requires not only extracting informative unimodal representations, but also modeling how temporally aligned behavioral evidence interacts across modalities. In this paper, we propose a synchronized visual-facial cross-refinement framework (SVF-CR) with pairwise multimodal evidence fusion for ambivalence and hesitancy recognition. The proposed method first extracts whole-video segment tokens and cropped-face segment tokens using the same temporal partition. The synchronized visual and facial tokens are refined through intra-modal self-attention and bidirectional visual-facial cross-attention, allowing whole-video context and local facial behavior to mutually refine each other before evidence construction. We then construct segment-level visual-facial evidence using consistency and discrepancy modeling, followed by temporal self-attention and attention pooling. Textual and acoustic features are lightly refined through context self-attention and are fused with the enhanced visual-facial evidence at the final decision stage using pairwise evidence fusion. Experiments on the BAH (Behavioral Ambivalence/Hesitancy) public evaluation split show that the proposed synchronized visual-facial cross-refinement improves public macro-F1 over both global visual-face token fusion and synchronized evidence baselines, achieving a public macro-F1 of 0.7156. Code is available at : https://github.com/hiinnnii/BAH-Challenge-ECCV2026\_SVF-CR.
Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades, indicating that models still struggle to identify and use the evidence most relevant to a question. A promising way to improve long-context utilization is test-time training (TTT), which treats the test context as a training example for instance-specific parameter adaptation. However, applying TTT to the entire long context is prohibitively expensive, while adapting on randomly sampled spans introduces severe noise. Because most spans in a long context are irrelevant to the specific question, training on them may even degrade the base model's performance. Our preliminary study shows that TTT is highly sensitive to training-span quality: on LongBench-v2, TTT on randomly sampled spans hurts performance, whereas TTT on oracle spans substantially improves it. Motivated by this, we propose a simple method, Self-Guided TTT (S-TTT): before adaptation, the model identifies the evidence spans it should learn from, and the standard language-modeling training objective is applied only to those selected spans. On two challenging long-context reasoning benchmarks, LongBench-v2 and LongBench-Pro, S-TTT improves accuracy for both Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, achieving up to a 15% relative improvement.
Worldbuilding, the construction of coherent fictional worlds, is a foundational task in game design and literary creation. Large Language Models (LLMs) offer new possibilities for automated content generation, but their application to worldbuilding faces three challenges: context explosion that grows linearly with the building process, the tension between creative diversity and content consistency, and the absence of automated quality assurance. This paper presents AutoWorldBuilder, a multi-agent collaborative system that addresses these challenges through five integrated components: a structured concept network with conflict detection; a DAG-based hybrid batch scheduler that groups tasks by semantic locality; a four-layer context compression mechanism achieving approximately 90% token reduction; an iterative review system with specialized Auditor agents that improves proposal pass rates from 42% to over 85%; and a skill-driven agent architecture supporting zero-code extension with differentiated temperature configuration. Two experiments across 20 diverse worldbuilding tasks, using GPT-OSS 120B and DeepSeek v3.2 as LLM backends, demonstrate a 95.0% success rate. The system generated 56-103 self-consistent concepts per world in 18-31 minutes with zero-conflict delivery. The architectural patterns validated here, including layer-as-budget compression, semantic-locality scheduling, and separation of generation and review, transfer to the broader class of knowledge-intensive, multi-agent LLM applications.
Adaptive power management in Electric Vehicles (EVs) requires accurate power prediction. Although deep learning models have emerged as highly effective for time-series forecasting in this domain, their performance is prone to degradation when exposed to data with distributions different from the training data. We introduce a novel approach that enables on-device learning in resource-constrained EV systems to continuously adapt pretrained battery prediction models to new, unseen data. We leverage existing pretrained models by transforming them into adaptable versions that retain critical hyperparameter knowledge from their initial training. We comprehensively investigate both online and offline model adaptation strategies. Our results demonstrate significant improvements in forecasting performance across various models and time horizons, achieving mean absolute error reductions of up to 7.49\% and 14.88\% with online and offline adaptation techniques, respectively. This study highlights the substantial benefit of on-device adaptation, resulting in enhanced battery power predictions than unadapted model deployments in real-world EV scenarios.
We introduce a novel method for both partial and full optimization of the connections in deep differentiable logic gate networks (LGNs) and lookup table networks (LUTNs). Our training method utilizes a probability distribution over a set of connections per gate/lookup table (LUT) input pin, selecting the connection with highest merit, all whilst the optimal gate types or LUT-entries are learned in parallel. We show that the connection-optimized LGNs outperform standard fixed-connection LGNs on the Yin-Yang, MNIST Handwritten Digits and Fashion-MNIST benchmarks, while requiring only a fraction of the number of logic gates. We achieve 98.92% on the MNIST dataset with two layers of 8000 gates. With only one layer of 8000 gates, we obtain 98.45%, showing that our method requires almost 50 times fewer gates compared to fixed-connection LGNs. Training stability up to ten layers has been ensured by employing a high learning rate, straight-through estimators and trimming constant-output gate types. Additionally, we present a LUT neuron description that enables stable training with backpropagation, tested up to 6-layer deep networks. The model requires four times fewer trainable parameters and still achieves a higher accuracy compared to the fixed-connection LGN training algorithm. Our connection-training algorithm also works well for the LUTNs, achieving an accuracy of 98.88% for two layers of 2000 6-input LUTs.
The growing adoption of large language model-based agents within operating system workflows has increased the importance of energy-efficient inference on laptop-class systems-on-chip (SoCs). While cloud offloading remains common, it introduces reliability and privacy concerns that are particularly problematic for agentic workloads. Recent laptop SoCs, therefore, incorporate neural processing engines (NPUs) optimized for energy efficiency; however, effectively mapping attention mechanisms onto NPUs remains challenging due to architectural diversity and explicit data-movement programming models. In this work, we present STEEL, the first open-source implementation of FlashAttention targeting XDNA-like NPUs. STEEL introduces a dataflow formulation of prefill attention, enabling efficient exploitation of spatial parallelism and on-chip memory. Furthermore, STEEL addresses the load imbalance induced by the causal mask by leveraging a sparsity-aware pipeline placement onto the NPU array, reducing synchronization overhead and improving utilization. We evaluate STEEL on the AMD Ryzen AI 9 HX 370 SoC and compare its performance against optimized CPU and GPU implementations. Experimental results show that STEEL reduces energy consumption by an average of 9.17x and 1.75x relative to CPU and GPU baselines, respectively. On XDNA 1, STEEL achieves an average 9.6x latency reduction over the prior state of the art, and delivers a 22.8x speedup on average compared to a layer-by-layer attention implementation on XDNA 2.
We study an adversarial bandit problem for entanglement-based quantum-network routing over a modest graph corpus. Alice selects an end-to-end repeater route for an Ekert-91 protocol (E91) representing her move, while Eve selects an attack surface, either edge intercept--resend or repeater memory degradation. Payoffs are drawn from cached SeQUeNCe-simulated E91 transcripts, and Alice accepts a turn when the finite-sample statistic violates the Clauser-Horne-Shimony-Holt (CHSH) bound. Performing adversarial co-learning across 50 structured topologies, we find that learned retention tracks a full-matrix minimax reference closely (Pearson $r=0.99$): under a one-surface Eve action model, bottleneck families have zero retention, while non-bottleneck families follow a $1-1/N$ coverage principle. We then fit decision-tree explanation models to graph-, attack-, and route-level topology-corpus targets and report their faithfulness. Finally, we construct prompt records for local language models to summarize the tree evidence, resulting in an open-source explanation workflow for quantum-repeater network games.
We present Mach-Mind-4-Flash, a 35B-parameter Mixture-of-Experts (MoE) agentic model with 3B activated parameters. Through post-training optimization alone without scaling pre-training compute, the model achieves performance on par with or surpassing that of 100B-parameter-class models. By introducing scalable agentic interaction environments for large-scale reinforcement learning, the model attains significant performance gains on real-world application tasks. Our pipeline comprises three stages: (1) a unified RL/OPD training infrastructure with dynamic multi-teacher scheduling and operator-level acceleration, delivering 17\% end-to-end training speedup; (2) multiple domain-specific RL experts trained in parallel across Reasoning, General, and Agent tracks, then fused into a single generalist via Multi-Teacher On-Policy Distillation (MOPD) -- a routed reverse-KL objective that eliminates the see-saw degradation of mixed-reward RL; (3) Hybrid Median-length Policy Optimization (HMPO), a single-stage token-efficiency method that compresses reasoning chains by 19--46\% with $\le$0.7 percentage-point accuracy loss. Mach-Mind-4-Flash scores 92.70 on AIME'26, 82.82 on IFBench, 80.74 on Behavioral-SafetyBench, 75.80 on BFCL-v4, 72.31 on BrowseComp-zh, and 84.20 on ClawBench -- leading or matching models with 10--30$\times$ its activated size at a fraction of the inference cost.
Program verification is crucial for software correctness, but producing fully verified programs remains difficult in practice. This paper studies whether implementation structure affects automated verifiability when multiple generated programs are intended to satisfy the same task-level semantics. We present Diversify2Verify, a staged LLM-based pipeline for Why3 that infers representation-specific contracts, generates and tests diverse recursive and imperative array/list implementations, and attempts verification with bounded verifier-guided annotation repair. We also construct a verification-oriented benchmark of 73 tasks over integers, arrays, and lists, yielding 292 implementation variants. Diversify2Verify verifies 96 artifacts initially and 154 after two repair passes, improving artifact-level verification from 32.9% to 52.7%. At the task level, at least one variant verifies for 49 of 73 tasks, a 67.1% success rate. These results show that task-equivalent implementations can differ substantially in verifiability and that implementation diversity helps find verification-friendly artifacts.