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 Hard2V erify, a human-annotated, step-level verification benchmark produced with over 500 hours of human labor. Hard2V erify is designed to rigorously assess step-level verifiers at the frontier: V erifiers 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.Figure 1: Comparison of models evaluated on both ProcessBench (Zheng et al., 2024a) and our Hard2V erify benchmark. Past benchmarks do not sufficiently evaluate in the frontier-level math settings that Hard2V erify does; On the same error identification task, Qwen2.5-Math-PRM-72B Mathematical reasoning serves as a gold-standard evaluation setting for benchmarking reasoning progress in large language models (LLMs). Over the past half-decade, benchmarks have been introduced to assess LLMs at the grade-school (Cobbe et al., 2021), high-school (Hendrycks et al., 2021), university (Zhang et al., 2023), and competition math level (MMA, 2025; He et al., 2024a; Gao et al., 2024). However, the progress of mathematical reasoning ability of LLMs has outpaced benchmark creation, with every subsequent release of a frontier LLM saturating new benchmarks, most recently with GPT -5 Pro achieving 96.5%+ on AIME 2024. As a result, recent efforts (Glazer et al., 2024; Phan et al., 2025) have written novel, unseen mathematical questions to test LLMs. 1 This paradigm requires training data with solutions that are easily verifiable, i.e., have solutions that can be easily checked against a known ground-truth by string matching or symbolic checkers. Math benchmarks, for the most part, also adopt the verifiable setup, where a model response is considered correct if its final answer matches the established ground-truth.

Generative AI systems are increasingly assisting and acting on behalf of end users in practical settings, from digital shopping assistants to next-generation autonomous cars. In this context, safety is no longer about blocking harmful content, but about preempting downstream hazards like financial or physical harm. Yet, most AI guardrails continue to rely on output classification based on labeled datasets and human-specified criteria,making them brittle to new hazardous situations. Even when unsafe conditions are flagged, this detection offers no path to recovery: typically, the AI system simply refuses to act--which is not always a safe choice. In this work, we argue that agentic AI safety is fundamentally a sequential decision problem: harmful outcomes arise from the AI system's continually evolving interactions and their downstream consequences on the world. We formalize this through the lens of safety-critical control theory, but within the AI model's latent representation of the world. This enables us to build predictive guardrails that (i) monitor an AI system's outputs (actions) in real time and (ii) proactively correct risky outputs to safe ones, all in a model-agnostic manner so the same guardrail can be wrapped around any AI model. We also offer a practical training recipe for computing such guardrails at scale via safety-critical reinforcement learning. Our experiments in simulated driving and e-commerce settings demonstrate that control-theoretic guardrails can reliably steer LLM agents clear of catastrophic outcomes (from collisions to bankruptcy) while preserving task performance, offering a principled dynamic alternative to today's flag-and-block guardrails.

Pruning is a common technique to reduce the compute and storage requirements of Neural Networks. While conventional approaches typically retrain the model to recover pruning-induced performance degradation, state-of-the-art Large Language Model (LLM) pruning methods operate layer-wise, minimizing the per-layer pruning error on a small calibration dataset to avoid full retraining, which is considered computationally prohibitive for LLMs. However, finding the optimal pruning mask is a hard combinatorial problem and solving it to optimality is intractable. Existing methods hence rely on greedy heuristics that ignore the weight interactions in the pruning objective. In this work, we instead consider the convex relaxation of these combinatorial constraints and solve the resulting problem using the Frank-Wolfe (FW) algorithm. Our method drastically reduces the per-layer pruning error, outperforms strong baselines on state-of-the-art GPT architectures, and remains memory-efficient. We provide theoretical justification by showing that, combined with the convergence guarantees of the FW algorithm, we obtain an approximate solution to the original combinatorial problem upon rounding the relaxed solution to integrality.

Despite the success of Reinforcement Learning from Human Feedback (RLHF) in aligning language models with human values, reward hacking-or reward over-optimization-remains a major challenge. We identify two key obstacles to its mitigation: (1) reward misgeneralization in reward modeling, where reward models overfit to spurious, preference-irrelevant features; and (2) the lack of suitable regularization during RL optimization, as existing token-level constraints often over-restrict the policy space. To address these issues, we propose InfoRM, an information-theoretic reward modeling framework based on the Information Bottleneck (IB) principle, which filters out preference-irrelevant information to alleviate reward misgeneralization. We further observe that reward-hacked responses manifest as pronounced outliers in InfoRM's IB latent space, measured by Mahalanobis distance from the SFT-induced distribution. Motivated by this, we introduce IBL, a distribution-level regularization that penalizes such deviations, effectively expanding the optimization landscape while maintaining alignment. We prove that IBL is theoretically equivalent to the pessimistic RL objective within the IB latent space. Finally, we present Mahalanobis Outlier Probability (MOP), a statistical metric for quantifying reward hacking severity, enabling principled hyperparameter tuning and online mitigation such as early stopping. Extensive experiments across diverse LLMs and datasets confirm the generality of our findings, the effectiveness of InfoRM and IBL, and the reliability of MOP as a diagnostic tool-collectively advancing the state of RLHF.

As texts generated by Large Language Models (LLMs) are ever more common and often indistinguishable from human-written content, research on automatic text detection has attracted growing attention. Many recent detectors report near-perfect accuracy, often boasting AUROC scores above 99\%. However, these claims typically assume fixed generation settings, leaving open the question of how robust such systems are to changes in decoding strategies. In this work, we systematically examine how sampling-based decoding impacts detectability, with a focus on how subtle variations in a model's (sub)word-level distribution affect detection performance. We find that even minor adjustments to decoding parameters - such as temperature, top-p, or nucleus sampling - can severely impair detector accuracy, with AUROC dropping from near-perfect levels to 1\% in some settings. Our findings expose critical blind spots in current detection methods and emphasize the need for more comprehensive evaluation protocols. To facilitate future research, we release a large-scale dataset encompassing 37 decoding configurations, along with our code and evaluation framework https://github.com/BaggerOfWords/Sampling-and-Detection

Large Language Model (LLM) inference has emerged as a fundamental paradigm. In real-world scenarios, variations in output length cause severe workload imbalance in the decode phase, particularly for long-output reasoning tasks. Existing systems, such as PD disaggregation architectures, rely on static prefill-to-decode scheduling, which often results in SLO violations and OOM failures under evolving decode workloads. In this paper, we propose ARES, an adaptive decoding rescheduling system powered by length prediction to anticipate future workloads. Our core contributions include: (1) A lightweight and continuous LLM-native prediction method that leverages LLM hidden state to model remaining generation length with high precision (reducing MAE by 49.42%) and low overhead (cutting predictor parameters by 93.28%); (2) A rescheduling solution in decode phase with : A dynamic balancing mechanism that integrates current and predicted workloads, reducing P99 TPOT by 74.77% and achieving up to 2.24 times higher goodput.

Time Series Foundation Models (TSFMs) represent a new paradigm for time series forecasting, offering zero-shot forecasting capabilities without the need for domain-specific pre-training or fine-tuning. However, as with Large Language Models (LLMs), evaluating TSFMs is tricky, as with ever more extensive training sets, it becomes more and more challenging to ensure the integrity of benchmarking data. Our investigation of existing TSFM evaluation highlights multiple challenges, ranging from the representativeness of the benchmark datasets, over the lack of spatiotemporal evaluation, to risks of information leakage due to overlapping and obscure datasets, and the memorization of global patterns caused by external shocks like economic crises or pandemics. Our findings reveal widespread confusion regarding data partitions, risking inflated performance estimates and incorrect transfer of global knowledge to local time series. We argue for the development of robust evaluation methodologies to prevent pitfalls already observed in LLM and classical time series benchmarking, and call upon the research community to design new, principled approaches, such as evaluations on truly out-of-sample future data, to safeguard the integrity of TSFM assessment.

Large Language Models (LLMs) can be adapted to extend their text capabilities to speech inputs. However, these speech-adapted LLMs consistently underperform their text-based counterparts--and even cascaded pipelines--on language understanding tasks. We term this shortfall the text-speech understanding gap: the performance drop observed when a speech-adapted LLM processes spoken inputs relative to when the original text-based LLM processes the equivalent text. Recent approaches to narrowing this gap either rely on large-scale speech synthesis of text corpora, which is costly and heavily dependent on synthetic data, or on large-scale proprietary speech datasets, which are not reproducible. As a result, there remains a need for more data-efficient alternatives for closing the text-speech understanding gap. In this work, we analyze the gap as driven by two factors: (i) forgetting of text capabilities during adaptation, and (ii) cross-modal misalignment between speech and text. Based on this analysis, we introduce SALAD--Sample-efficient Alignment with Learning through Active selection and cross-modal Distillation-- which combines cross-modal distillation with targeted synthetic data to improve alignment while mitigating forgetting. Applied to 3B and 7B LLMs, SALAD achieves competitive performance with a strong open-weight model across broad-domain benchmarks in knowledge, language understanding, and reasoning, while training on over an order of magnitude less speech data from public corpora.Figure 1: SALAD reduces the text-speech understanding gap while requiring over an order of magnitude less training data than competing speech-adapted LLMs. Work done during an internship at Apple. Large language models (LLMs) have demonstrated impressive capabilities in general knowledge and reasoning, often surpassing specialized systems across a wide range of tasks.

Accurate coding of tumor diagnoses with ICD-10-GM and ICD-O-3 is essential for structured cancer documentation in Germany. Smaller open-weight LLMs are appealing for privacy-preserving automation but often struggle with coding accuracy in German-language contexts. This study investigates whether instruction-based fine-tuning on public datasets improves the coding accuracy of open-weight LLMs for German tumor diagnosis texts. The evaluation uses coded diagnoses from the local tumor documentation system as test data. In a systematic data quality assessment, the upper limit for ICD-10 coding performance was estimated at 60-79% for exact and 81-94% for partial (three-character codes only) derivation. As training data, over 500,000 question-answer pairs were created based on the ICD-10-GM, ICD-O-3, and OPS catalogues. Eight open-weight models from the Qwen, Llama, and Mistral families (7-70 B parameters) were fine-tuned. ICD-10-GM accuracy rose from 1.4-24% to 41-58%, and partial accuracy from 31-74% to 73-83%. The accuracy of ICD-O-3 topography coding also improved but started and remained considerably lower with an exact accuracy of 22-40% and a partial accuracy of 56-67% after fine-tuning. Malformed code outputs dropped to 0% for all models. Tumor-diagnosis recognition reached 99%. Accuracy correlated positively with model size, but gaps between small and large models narrowed after fine-tuning. The reasoning mode in Qwen3 generally yielded a lower performance than fine-tuning and was over 100 times slower. Our findings highlight the potential of leveraging public catalogues to build instruction datasets that improve LLMs in medical documentation tasks. The complete training dataset and the best-performing checkpoints of the fine-tuned models are available from https://huggingface.co/datasets/stefan-m-lenz/ICDOPS-QA-2024.

Artificial Intelligence (AI) and machine learning (ML) models have made stark advances in the past three years, fueled by the development of foundation models (FM) trained on large-scale multimodal data. Following the public release of several successful FMs (OpenAI (2022); Brown et al. (2020); Bommasani et al. (2022)), FMs such as large language models (LLMs) and vision language models (VLMs) have bridged vision, language, and other modalities. In many Computer Science subfields such as Natural Language Processing (NLP) and Computer Vision (CV), FMs have demonstrated strong compositional performance and generalization capabilities (Awais et al. (2025); Gunter et al. (2024)), emerging as widely-used tools (Bommasani et al. (2022)) that provide a flexible backbones for innovation in other fields (Moor et al. (2023); Sartor & Thompson (2025); Firoozi et al. (2024)). Conducting FM research requires significant data, computing, and human resources (Cottier et al. (2024); Maslej et al. (2024); Crawford (2024)). A central concern in the field is whether greater access to such resources directly translates into more impactful research outcomes (Acemoglu (2024); Dodge et al. (2019); OpenAI (2018)), such as more research publications, or higher citation counts (Sinclair et al. (2023); Anjum et al. (2019)). The answer to this question has important implications for how resources are allocated, which research directions are prioritized, and how equitable participation in FM research can be ensured. However, the cost of research is often difficult to quantify due to lack of uniform disclosure on resource distribution (Bommasani et al. (2024)). Absent widespread disclosure, funding is perhaps most easily characterized in the concrete cost of purchasing or renting hardware (e.g., computing clusters, or chips), through there are also software, cloud storage services, and specialized software platform costs.