Conformal Certification of Reasoning Trace Prefixes

Language model reasoning traces are rarely all-or-nothing; they frequently contain valid intermediate steps before a critical error occurs. Existing uncertainty quantification methods typically certify final answers or entire responses, failing to provide statistical guarantees for the proportion of a sequential trace that can be safely retained. To address this, we introduce CROP (Conformal Reasoning Output Prefixes), a verifier-agnostic calibration procedure for clean-prefix certification. Given any step-level risk proxy, CROP selects a calibrated threshold and returns the longest contiguous prefix whose step risk proxies remain below it, routing the uncertified suffix for downstream review or repair. Assuming exchangeability, CROP rigorously controls the marginal probability that the returned prefix contains an annotated error. Across six process-labeled reasoning datasets, we demonstrate that standard step-level metrics such as AUROC do not fully capture prefix utility, suggesting verifiers should instead be evaluated by certified prefix length. Furthermore, CROP balances over- and under-withholding, improving downstream repair accuracy by preserving valid intermediate reasoning while discarding misleading suffixes. Ultimately, this work positions prefix certification as a rigorous, practical bridge between process supervision, abstention, and repair.

LT2: Linear-Time Looped Transformers

Looped Transformers (LT) have emerged as a powerful architecture by iterating their layers multiple times before decoding the final token. However, pairing them with full attention retains quadratic complexity, making them computationally expensive and slow. We introduce LT2 (Linear-Time Looped Transformers), a family of looped architectures that replace quadratic softmax attention with subquadratic, linear-time attention. We study two variants: LT2-linear with linear attention and LT2-sparse with sparse attention. We find that looping uniquely synergizes with these variants: it enables iterative memory refinement in linear attention and progressively expands the effective receptive field in sparse attention. We formalize these benefits theoretically and demonstrate consistent empirical gains across controlled recall, state-tracking, and language modeling tasks. We then explore LT2-hybrid, which combines different attention variants in a looped setting. Two variants are especially promising: LT2-hybrid (GDN+DSA), which interleaves linear and sparse attention to maximize efficiency and matches the standard looped transformer's quality at fully linear-time cost; and LT2-hybrid (Full+GDN), which interleaves GDN with a small fraction of full attention layers to maximize quality, surpassing the standard looped transformer in both performance and efficiency. We also show how to convert a pre-trained LT into an LT2-hybrid model. With about 1B tokens of training, our converted model, Ouro-hybrid-1.4B, outperforms industry-level 1B models and is competitive with industry-level 4B models while retaining the speed benefits of linear-time attention. Together, these results show a clear path toward making looped transformers more scalable and advancing efficient, capable small language models.

BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD

Industrial Computer-Aided Design (CAD) code generation requires models to produce executable parametric programs from visual or textual inputs. Beyond recognizing the outer shape of a part, this task involves understanding its 3D structure, inferring engineering parameters, and choosing CAD operations that reflect how the part would be designed and manufactured. Despite the promise of Multimodal large language models (MLLMs) for this task, they are rarely evaluated on whether these capabilities jointly hold in realistic industrial CAD settings. We present BenchCAD, a unified benchmark for industrial CAD reasoning. BenchCAD contains 17,900 execution-verified CadQuery programs across 106 industrial part families, including bevel gears, compression springs, twist drills, and other reusable engineering designs. It evaluates models through visual question answering, code question answering, image-to-code generation, and instruction-guided code editing, enabling fine-grained analysis across perception, parametric abstraction, and executable program synthesis. Across 10+ frontier models, BenchCAD shows that current systems often recover coarse outer geometry but fail to produce faithful parametric CAD programs. Common failures include missing fine 3D structure, misinterpreting industrial design parameters, and replacing essential operations such as sweeps, lofts, and twist-extrudes with simpler sketch-and-extrude patterns. Fine-tuning and reinforcement learning improve in-distribution performance, but generalization to unseen part families remains limited. These results position BenchCAD as a benchmark for measuring and improving the industrial readiness of multimodal CAD automation.

Agentic Discovery with Active Hypothesis Exploration for Visual Recognition

We introduce HypoExplore, an agentic framework that formulates neural architecture discovery for visual recognition as a hypothesis-driven scientific inquiry. Given a human-specified high-level research direction, HypoExplore ideates, implements, evaluates, and improves neural architectures through evolutionary branching. New hypotheses are created using a large language model by selecting a parent hypothesis to build upon, guided by a dual strategy that balances exploiting validated principles with resolving uncertain ones. Our proposed framework maintains a Trajectory Tree that records the lineage of all proposed architectures, and a Hypothesis Memory Bank that actively tracks confidence scores acquired through experimental evidence. After each experiment, multiple feedback agents analyze the results from different perspectives and consolidate their findings into hypothesis confidence updates. Our framework is tested on discovering lightweight vision architectures on CIFAR-10, with the best achieving 94.11% accuracy evolved from a root node baseline that starts at 18.91%, and generalizes to CIFAR-100 and Tiny-ImageNet. We further demonstrate applicability to a specialized domain by conducting independent architecture discovery runs on MedMNIST, which yield a state-of-the-art performance. We show that hypothesis confidence scores grow increasingly predictive as evidence accumulates, and that the learned principles transfer across independent evolutionary lineages, suggesting that HypoExplore not only discovers stronger architectures, but can help build a genuine understanding of the design space.

PRISM: A Dual View of LLM Reasoning through Semantic Flow and Latent Computation

Large language models (LLMs) solve complex problems by generating multi-step reasoning traces. Yet these traces are typically analyzed from only one of two perspectives: the sequence of tokens across different reasoning steps in the generated text, or the hidden-state vectors across model layers within one step. We introduce PRISM (Probabilistic Reasoning Inspection through Semantic and Implicit Modeling), a framework and diagnostic tool for jointly analyzing both levels, providing a unified view of how reasoning evolves across steps and layers. Across multiple reasoning models and benchmarks, PRISM uncovers systematic patterns in the reasoning process, showing that failed trajectories are more likely to become trapped in unproductive verification loops and further diverge into distinct modes such as overthinking and premature commitment, which behave differently once a candidate answer is reached. It further reveals how prompting reshapes reasoning behavior beyond aggregate accuracy by altering both semantic transitions and internal computational patterns. By modeling reasoning trajectories as structured processes, PRISM makes these behaviors observable and analyzable rather than relying solely on final-task accuracy. Taken together, these insights position PRISM as a practical tool for analyzing and diagnosing reasoning processes in LLMs.

GR-SAP: Generative Replay for Safety Alignment Preservation during Fine-Tuning

Recent studies show that the safety alignment of large language models (LLMs) can be easily compromised even by seemingly non-adversarial fine-tuning. To preserve safety alignment during fine-tuning, a widely used strategy is to jointly optimize safety and task objectives by mixing in the original alignment data, which is typically inaccessible even for open-weight LLMs. Inspired by generative replay in continual learning, we propose Generative Replay for Safety Alignment Preservation (GR-SAP), a unified framework that synthesizes domain-specific alignment data from LLMs and integrate them during downstream adaption to preserve safety alignment. Theoretical and empirical analyses demonstrate this synthetic data serves as a reliable proxy for the original alignment data. Experiments across various models and downstream tasks show that GR-SAP substantially mitigates fine-tuning-induced safety degradation while maintaining comparable downstream performance. Our code is available at https://github.com/chili-lab/gr-sap.

Spherical Steering: Geometry-Aware Activation Rotation for Language Models

Inference-time steering has emerged as a promising paradigm for controlling language models (LMs) without the cost of retraining. However, standard approaches typically rely on activation addition, a geometric operation that inevitably alters the magnitude of hidden representations. This raises concerns about representation collapse and degradation of open-ended generation capabilities. In this work, we explore Spherical Steering, a training-free primitive that resolves this trade-off through activation rotation. Rather than shifting activations with a fixed vector, our method rotates them along a geodesic toward a target direction, guiding the activation toward the target concept while preserving the integrity of the signal. To further enhance adaptivity, we incorporate a confidence gate that dynamically modulates steering strength based on input uncertainty. Extensive experiments across multiple-choice benchmarks demonstrate that Spherical Steering significantly outperforms addition-based baselines (notably by +10% on TruthfulQA, COPA, and Storycloze), while simultaneously maintaining the model's general open-ended generation quality. This work highlights the value of geometric consistency, suggesting that norm-preserving rotation is a robust and effective primitive for precise inference-time control.

SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports

Deeply understanding sports requires an intricate blend of fine-grained visual perception and rule-based reasoning - a challenge that pushes the limits of current multimodal models. To succeed, models must master three critical capabilities: perceiving nuanced visual details, applying abstract sport rule knowledge, and grounding that knowledge in specific visual evidence. Current sports benchmarks either cover single sports or lack the detailed reasoning chains and precise visual grounding needed to robustly evaluate these core capabilities in a multi-sport context. To address this gap, we introduce SportR, the first multi-sports large-scale benchmark designed to train and evaluate MLLMs on the fundamental reasoning required for sports intelligence. Our benchmark provides a dataset of 5,017 images and 2,101 videos. To enable granular evaluation, we structure our benchmark around a progressive hierarchy of question-answer (QA) pairs designed to probe reasoning at increasing depths - from simple infraction identification to complex penalty prediction. For the most advanced tasks requiring multi-step reasoning, such as determining penalties or explaining tactics, we provide 7,118 high-quality, human-authored Chain of Thought (CoT) annotations. In addition, our benchmark incorporates both image and video modalities and provides manual bounding box annotations to test visual grounding in the image part directly. Extensive experiments demonstrate the profound difficulty of our benchmark. State-of-the-art baseline models perform poorly on our most challenging tasks. While training on our data via Supervised Fine-Tuning and Reinforcement Learning improves these scores, they remain relatively low, highlighting a significant gap in current model capabilities. SportR presents a new challenge for the community, providing a critical resource to drive future research in multimodal sports reasoning.

DeepSport: A Multimodal Large Language Model for Comprehensive Sports Video Reasoning via Agentic Reinforcement Learning

Sports video understanding presents unique challenges, requiring models to perceive high-speed dynamics, comprehend complex rules, and reason over long temporal contexts. While Multimodal Large Language Models (MLLMs) have shown promise in genral domains, the current state of research in sports remains narrowly focused: existing approaches are either single-sport centric, limited to specific tasks, or rely on training-free paradigms that lack robust, learned reasoning process. To address this gap, we introduce DeepSport, the first end-to-end trained MLLM framework designed for multi-task, multi-sport video understanding. DeepSport shifts the paradigm from passive frame processing to active, iterative reasoning, empowering the model to ``think with videos'' by dynamically interrogating content via a specialized frame-extraction tool. To enable this, we propose a data distillation pipeline that synthesizes high-quality Chain-of-Thought (CoT) trajectories from 10 diverse data source, creating a unified resource of 78k training data. We then employ a two-stage training strategy, Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) with a novel gated tool-use reward, to optimize the model's reasoning process. Extensive experiments on the testing benchmark of 6.7k questions demonstrate that DeepSport achieves state-of-the-art performance, significantly outperforming baselines of both proprietary model and open-source models. Our work establishes a new foundation for domain-specific video reasoning to address the complexities of diverse sports.

Learning Distribution-Wise Control in Representation Space for Language Models

Interventions in language models (LMs) are applied strategically to steer model behavior during the forward pass. Learnable interventions, also known as representation fine-tuning, aim to apply pointwise control within the concept subspace and have proven effective in altering high-level behaviors. In this work, we extend this approach to the distribution level, enabling the model to learn not only pointwise transformations but also the surrounding regions of the concept subspace. We demonstrate that these methods perform effectively in early layers, with larger standard deviations correlating strongly with improved performance. Across eight commonsense reasoning and seven arithmetic reasoning benchmarks, our distribution-wise interventions consistently outperform pointwise interventions in controllability and robustness. These results illustrate that distribution-wise interventions provide a more comprehensive method for steering model behavior and enabling finer-grained control over language models. The code is at: \href{https://github.com/chili-lab/D-Intervention}{https://github.com/chili-lab/D-Intervention}.