This manuscript investigates the integration of positional encoding -- a technique widely used in computer graphics -- into the input vector of a binary classification model for self-collision detection. The results demonstrate the benefits of incorporating positional encoding, which enhances classification accuracy by enabling the model to better capture high-frequency variations, leading to a more detailed and precise representation of complex collision patterns. The manuscript shows that machine learning-based techniques, such as lightweight multilayer perceptrons (MLPs) operating in a low-dimensional feature space, offer a faster alternative for collision checking than traditional methods that rely on geometric approaches, such as triangle-to-triangle intersection tests and Bounding Volume Hierarchies (BVH) for mesh-based models.

While current onboard state estimation methods are adequate for most driving and safety-related applications, they do not provide insights into the interaction between tires and road surfaces. This paper explores a novel communication concept for efficiently transmitting integrated wheel sensor data from an ESP32 microcontroller. Our proposed approach utilizes a publish-subscribe system, surpassing comparable solutions in the literature regarding data transmission volume. We tested this approach on a drum tire test rig with our prototype sensors system utilizing a diverse selection of sample frequencies between 1 Hz and 32 000 Hz to demonstrate the efficacy of our communication concept. The implemented prototype sensor showcases minimal data loss, approximately 0.1% of the sampled data, validating the reliability of our developed communication system. This work contributes to advancing real-time data acquisition, providing insights into optimizing integrated wheel sensor communication.

The long-standing division between \textit{online} and \textit{offline} Multi-Object Tracking (MOT) has led to fragmented solutions that fail to address the flexible temporal requirements of real-world deployment scenarios. Current \textit{online} trackers rely on frame-by-frame hand-crafted association strategies and struggle with long-term occlusions, whereas \textit{offline} approaches can cover larger time gaps, but still rely on heuristic stitching for arbitrarily long sequences. In this paper, we introduce NOOUGAT, the first tracker designed to operate with arbitrary temporal horizons. NOOUGAT leverages a unified Graph Neural Network (GNN) framework that processes non-overlapping subclips, and fuses them through a novel Autoregressive Long-term Tracking (ALT) layer. The subclip size controls the trade-off between latency and temporal context, enabling a wide range of deployment scenarios, from frame-by-frame to batch processing. NOOUGAT achieves state-of-the-art performance across both tracking regimes, improving \textit{online} AssA by +2.3 on DanceTrack, +9.2 on SportsMOT, and +5.0 on MOT20, with even greater gains in \textit{offline} mode.

Recent advancements highlight the importance of GRPO-based reinforcement learning methods and benchmarking in enhancing text-to-image (T2I) generation. However, current methods using pointwise reward models (RM) for scoring generated images are susceptible to reward hacking. We reveal that this happens when minimal score differences between images are amplified after normalization, creating illusory advantages that drive the model to over-optimize for trivial gains, ultimately destabilizing the image generation process. To address this, we propose Pref-GRPO, a pairwise preference reward-based GRPO method that shifts the optimization objective from score maximization to preference fitting, ensuring more stable training. In Pref-GRPO, images are pairwise compared within each group using preference RM, and the win rate is used as the reward signal. Extensive experiments demonstrate that PREF-GRPO differentiates subtle image quality differences, providing more stable advantages and mitigating reward hacking. Additionally, existing T2I benchmarks are limited by coarse evaluation criteria, hindering comprehensive model assessment. To solve this, we introduce UniGenBench, a unified T2I benchmark comprising 600 prompts across 5 main themes and 20 subthemes. It evaluates semantic consistency through 10 primary and 27 sub-criteria, leveraging MLLM for benchmark construction and evaluation. Our benchmarks uncover the strengths and weaknesses of both open and closed-source T2I models and validate the effectiveness of Pref-GRPO.

Autonomous planetary exploration demands real-time, high-fidelity environmental perception. Standard deep learning models require massive computational resources. Conversely, space-qualified onboard computers operate under strict power, thermal, and memory limits. This disparity creates a severe engineering bottleneck, preventing the deployment of highly capable perception architectures on extraterrestrial exploration platforms. In this foundational concept paper, we propose the theoretical architecture for the Adaptive Quantized Planetary Crater Detection System (AQ-PCDSys) to resolve this bottleneck. We present a mathematical blueprint integrating an INT8 Quantized Neural Network (QNN) designed specifically for Quantization Aware Training (QAT). To address sensor fragility, we mathematically formalize an Adaptive Multi-Sensor Fusion (AMF) module. By deriving the exact integer requantization multiplier required for spatial attention gating, this module actively selects and fuses Optical Imagery (OI) and Digital Elevation Models (DEMs) at the feature level, ensuring reliable perception during extreme cross-illuminations and optical hardware dropouts. Furthermore, the architecture introduces anchor-free, center-to-edge regression heads, protected by a localized FP16 coordinate conversion, to accurately frame asymmetrical lunar craters without catastrophic integer truncation. Rather than presenting physical hardware telemetry, this manuscript establishes the theoretical bounds, structural logic, and mathematical justifications for the architecture. We outline a rigorous Hardware-in-the-Loop (HITL) evaluation protocol to define the exact testing criteria required for future empirical validation, paving the way for next-generation space-mission software design.

Large language models (LLMs) have demonstrated significant capabilities in solving mathematical problems expressed in natural language. However, multilingual and culturally-grounded mathematical reasoning in low-resource languages lags behind English due to the scarcity of socio-cultural task datasets that reflect accurate native entities such as person names, organization names, and currencies. Existing multilingual benchmarks are predominantly produced via translation and typically retain English-centric entities, owing to the high cost associated with human annotater-based localization. Moreover, automated localization tools are limited, and hence, truly localized datasets remain scarce. To bridge this gap, we introduce a framework for LLM-driven cultural localization of math word problems that automatically constructs datasets with native names, organizations, and currencies from existing sources. We find that translated benchmarks can obscure true multilingual math ability under appropriate socio-cultural contexts. Through extensive experiments, we also show that our framework can help mitigate English-centric entity bias and improves robustness when native entities are introduced across various languages.

LLMs increasingly exhibit over-refusal behavior, where safety mechanisms cause models to reject benign instructions that seemingly resemble harmful content. This phenomenon diminishes utility in production applications that repeatedly rely on common prompt templates or applications that frequently rely on LLMs for specific tasks (e.g. sentiment analysis, language translation). Through extensive evaluation, we demonstrate that LLMs persist in refusing inputs containing harmful content, even when they are reframed with tasks that have benign intent. Our mechanistic analysis reveals that LLMs follow distinct "constellation" patterns in embedding space as representations traverse layers, with each NLP task maintaining consistent trajectories that shift predictably between refusal and non-refusal cases. We introduce SafeConstellations, an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways. By selectively guiding model behavior only on tasks prone to over-refusal, our method reduces over-refusals with minimal impact on utility -- offering a principled and conditional approach to mitigating over-refusals.

Modifications to test-time sampling have emerged as an important extension to diffusion algorithms, with the goal of biasing the generative process to achieve a given objective without having to retrain the entire diffusion model. However, generating jointly correlated samples from multiple pre-trained diffusion models while simultaneously enforcing task-specific constraints without costly retraining has remained challenging. To this end, we propose Projected Coupled Diffusion (PCD), a novel test-time framework for constrained joint generation. PCD introduces a coupled guidance term into the generative dynamics to encourage coordination between diffusion models and incorporates a projection step at each diffusion step to enforce hard constraints. Empirically, we demonstrate the effectiveness of PCD in application scenarios of image-pair generation, object manipulation, and multi-robot motion planning. Our results show improved coupling effects and guaranteed constraint satisfaction without incurring excessive computational costs.

Reinforcement learning with verifiable reward (RLVR) has become a promising paradigm for post-training large language models (LLMs) to improve their reasoning capability. However, when the rollout accuracy is low on hard problems, the reward becomes sparse, limiting learning efficiency and causing exploration bottlenecks. Existing approaches either rely on teacher models for distillation or filter out difficult problems, which limits scalability or restricts reasoning improvement through exploration. We propose EvoCoT, a self-evolving curriculum learning framework based on two-stage chain-of-thought (CoT) reasoning optimization. EvoCoT constrains the exploration space by self-generating and verifying CoT trajectories, then gradually shortens CoT steps to expand the space in a controlled way. The framework enables LLMs to stably learn from initially unsolved hard problems under sparse rewards. We apply EvoCoT to multiple LLM families, including Qwen, DeepSeek, and Llama. Experiments show that EvoCoT enables LLMs to solve previously unsolved problems, improves reasoning capability without external CoT supervision, and is compatible with various RL fine-tuning methods. We release the source code to support future research.

As medical LLMs transition to clinical deployment, assessing their ethical reasoning capability becomes critical. While achieving high accuracy on knowledge benchmarks, LLMs lack validated assessment for navigating ethical trade-offs in clinical decision-making where multiple valid solutions exist. Existing benchmarks lack systematic approaches to incorporate recognized philosophical frameworks and expert validation for ethical reasoning assessment. We introduce PrinciplismQA, a philosophy-grounded approach to assessing LLM clinical medical ethics alignment. Grounded in Principlism, our approach provides a systematic methodology for incorporating clinical ethics philosophy into LLM assessment design. PrinciplismQA comprises 3,648 expert-validated questions spanning knowledge assessment and clinical reasoning. Our expert-calibrated pipeline enables reproducible evaluation and models ethical biases. Evaluating recent models reveals significant ethical reasoning gaps despite high knowledge accuracy, demonstrating that knowledge-oriented training does not ensure clinical ethical alignment. PrinciplismQA provides a validated tool for assessing clinical AI deployment readiness.

Robotic task planning in real-world environments requires reasoning over implicit constraints from language and vision. While LLMs and VLMs offer strong priors, they struggle with long-horizon structure and symbolic grounding. Existing methods that combine LLMs with symbolic planning often rely on handcrafted or narrow domains, limiting generalization. We propose UniDomain, a framework that pre-trains a PDDL domain from robot manipulation demonstrations and applies it for online robotic task planning. It extracts atomic domains from 12,393 manipulation videos to form a unified domain with 3137 operators, 2875 predicates, and 16481 causal edges. Given a target class of tasks, it retrieves relevant atomics from the unified domain and systematically fuses them into high-quality meta-domains to support compositional generalization in planning. Experiments on diverse real-world tasks show that UniDomain solves complex, unseen tasks in a zero-shot manner, achieving up to 58% higher task success and 160% improvement in plan optimality over state-of-the-art LLM and LLM-PDDL baselines.

We study how to learn treatment policies from multimodal electronic health records (EHRs) that consist of tabular data and clinical text. These policies can help physicians make better treatment decisions and allocate healthcare resources more efficiently. Causal policy learning methods prioritize patients with the largest expected treatment benefit. Yet, existing estimators are designed for tabular covariates under causal assumptions that may be hard to justify in the multimodal setting. A pragmatic alternative is to apply causal estimators directly to multimodal representations, but this can produce biased treatment effect estimates when the representations do not preserve the relevant confounding information. As a result, predictive models of baseline risk are commonly used in practice to guide treatment decisions, although they are not designed to identify which patients benefit most from treatment. We propose AACE (Annotation-Assisted Coarsened Effects), an annotation-assisted approach to causal policy learning for multimodal EHRs. The method uses expert-provided annotations during training to support confounding adjustment, and then predicts treatment benefit from only multimodal representations at inference. We show that the proposed method achieves strong empirical performance across synthetic, semi-synthetic, and real-world EHR datasets, outperforming risk-based and representation-based causal baselines, and offering practical insights for applying causal machine learning in clinical practice.

Retrieval-Augmented Generation (RAG) effectively improves the accuracy of Large Language Models (LLMs). However, retrieval noises significantly undermine the quality of LLMs' generation, necessitating the development of denoising mechanisms. Previous works extract evidence straightforwardly without deep thinking, which may risk filtering out key clues and struggle with generalization. To this end, we propose EviOmni, which learns to extract rational evidence via reasoning first and then extracting. Specifically, EviOmni integrates evidence reasoning and evidence extraction into one unified trajectory, followed by knowledge token masking to avoid information leakage, optimized via on-policy reinforcement learning with verifiable rewards in terms of answer, length, and format. Extensive experiments on five benchmark datasets show the superiority of EviOmni, which provides compact and high-quality evidence, enhances the accuracy of downstream tasks, and supports both traditional and agentic RAG systems.

Vision-language models (VLMs) increasingly combine visual and textual information to perform complex tasks. However, conflicts between their internal knowledge and external visual input can lead to hallucinations and unreliable predictions. In this work, we investigate the mechanisms that VLMs use to resolve cross-modal conflicts by introducing WHOOPS-AHA!, a dataset of multimodal counterfactual queries that deliberately contradict internal commonsense knowledge. Through logit inspection, we identify a small set of attention heads that mediate this conflict. By intervening in these heads, we can steer the model towards its internal parametric knowledge or the visual information. Our results show that attention patterns on these heads effectively locate image regions that influence visual overrides, providing a more precise attribution compared to gradient-based methods.

We describe a vulnerability in language models (LMs) trained with user feedback, whereby a single user can persistently alter LM knowledge and behavior given only the ability to provide prompts and upvote / downvote feedback on LM outputs. To implement the attack, the attacker prompts the LM to stochastically output either a "poisoned" or benign response, then upvotes the poisoned response or downvotes the benign one. When feedback signals are used in a subsequent preference tuning behavior, LMs exhibit increased probability of producing poisoned responses even in contexts without malicious prompts. We show that this attack can be used to (1) insert factual knowledge the model did not previously possess, (2) modify code generation patterns in ways that introduce exploitable security flaws, and (3) inject fake financial news. Our finding both identifies a new qualitative feature of language model preference tuning (showing that it even highly restricted forms of preference data can be used to exert fine-grained control over behavior), and a new attack mechanism for LMs trained with user feedback (extending work on pretraining-time data poisoning and deployment-time prompt injection).

We show that a language model's ability to predict text is tightly linked to the breadth of its embedding space: models that spread their contextual representations more widely tend to achieve lower perplexity. Concretely, we find that representation dispersion--the average pairwise cosine distance among hidden vectors--strongly and negatively correlates with perplexity across diverse model families (LLaMA, Qwen, and others) and domains (Wikipedia, news, scientific abstracts). Beyond illustrating this link, we show how dispersion can be leveraged for a range of practical tasks--without requiring labeled data. First, measuring dispersion on unlabeled text allows us to rank examples by difficulty and identify hard slices in new domains, offering a data-efficient tool for screening and prioritizing models before full evaluation. Next, we find that identifying layers with higher dispersion pinpoints the best representations for retrieval-based methods such as kNN-LM, bypassing exhaustive layer-by-layer searches. Finally, we integrate a simple "push-away" objective into training, which increases dispersion in both single-domain and cross-domain scenarios and directly improves perplexity in each. Code is available at https://github.com/yanhong-lbh/rep_dispersion.

Negation is a fundamental linguistic phenomenon that poses ongoing challenges for Large Language Models (LLMs), particularly in tasks requiring deep semantic understanding. Current benchmarks often treat negation as a minor detail within broader tasks, such as natural language inference. Consequently, there is a lack of benchmarks specifically designed to evaluate comprehension of negation. In this work, we introduce Thunder-NUBench, a novel benchmark explicitly created to assess sentence-level understanding of negation in LLMs. Thunder-NUBench goes beyond merely identifying surface-level cues by contrasting standard negation with structurally diverse alternatives, such as local negation, contradiction, and paraphrase. This benchmark includes manually curated sentence-negation pairs and a multiple-choice dataset, allowing for a comprehensive evaluation of models' understanding of negation.

While Mamba has demonstrated strong performance in language modeling, its potential as a speech self-supervised learning (SSL) model remains underexplored, with prior studies limited to isolated tasks. To address this, we explore Mamba-based HuBERT models as alternatives to Transformer-based SSL architectures. Leveraging the linear-time Selective State Space, these models enable fine-tuning on long-context ASR with significantly lower compute. Moreover, they show superior performance when fine-tuned for streaming ASR. Beyond fine-tuning, these models show competitive performance on SUPERB probing benchmarks, particularly in causal settings. Our analysis shows that they yield higher-quality quantized representations and capture speaker-related features more distinctly than Transformer-based models. These findings highlight Mamba-based SSL as a promising and complementary direction for long-sequence modeling, real-time speech modeling, and speech unit extraction. The codebase is available at https://github.com/hckuo145/Mamba-based-HuBERT.

Multi-task learning for dense prediction is limited by the need for extensive annotation for every task, though recent works have explored training with partial task labels. Leveraging the generalization power of diffusion models, we extend the partial learning setup to a zero-shot setting, training a multi-task model on multiple synthetic datasets, each labeled for only a subset of tasks. Our method, StableMTL, repurposes image generators for latent regression. Adapting a denoising framework with task encoding, per-task conditioning and a tailored training scheme. Instead of per-task losses requiring careful balancing, a unified latent loss is adopted, enabling seamless scaling to more tasks. To encourage inter-task synergy, we introduce a multi-stream model with a task-attention mechanism that converts N-to-N task interactions into efficient 1-to-N attention, promoting effective cross-task sharing. StableMTL outperforms baselines on 7 tasks across 8 benchmarks.

Recent progress in large language models (LLMs) has boosted mathematical reasoning, yet geometry remains challenging where auxiliary construction is often essential. Prior methods either underperform or depend on very large models (e.g., GPT-4o), making them costly. We argue that reinforcement learning with verifiable rewards (e.g., GRPO) can train smaller models to couple auxiliary construction with solid geometric reasoning. However, naively applying GRPO yields unconditional rewards, encouraging indiscriminate and sometimes harmful constructions. We propose Group Contrastive Policy Optimization (GCPO), an RL framework with two components: (1) Group Contrastive Masking, which assigns positive/negative construction rewards based on contextual utility, and (2) a Length Reward that encourages longer reasoning chains. On top of GCPO, we build GeometryZero, an affordable family of geometry reasoning models that selectively use auxiliary construction. Experiments on Geometry3K and MathVista show GeometryZero consistently outperforms RL baselines (e.g., GRPO, ToRL). The code has been available at https://github.com/ekonwang/GeometryZero.

Building Task-Oriented Dialogue (TOD) systems that generalize across different tasks remains a challenging problem. Data-driven approaches often struggle to transfer effectively to unseen tasks. While recent schema-based TOD frameworks improve generalization by decoupling task logic from language understanding, their reliance on neural or generative models often obscures how task schemas influence behaviour and hence impair interpretability. In this work, we introduce a novel framework, CoDial (Code for Dialogue), at the core of which is converting a predefined task schema to a structured heterogeneous graph and then to programmatic LLM guardrailing code, such as NVIDIA's Colang. The pipeline enables efficient and interpretable alignment of dialogue policies during inference. We introduce two paradigms for LLM guardrailing code generation, $\text{CoDial}_{\text{free}}$ and $\text{CoDial}_{\text{structured}}$, and propose a mechanism that integrates human feedback to iteratively improve the generated code. Empirically, CoDial achieves state-of-the-art (SOTA) performance on the widely used benchmark datasets, while providing inherent interpretability in the design. We additionally demonstrate CoDial's iterative improvement via manual and LLM-aided feedback, making it a practical tool for human-guided alignment of LLMs in unseen domains.

Multimodal language models (MLMs) increasingly demonstrate human-like communication, yet their use of everyday perspectival words remains poorly understood. To address this gap, we compare humans and MLMs in their use of three word types that impose increasing cognitive demands: vocabulary (for example, "boat" or "cup"), possessives (for example, "mine" versus "yours"), and demonstratives (for example, "this one" versus "that one"). Testing seven MLMs against human participants, we find that perspectival words are harder than vocabulary words for both groups. The gap is larger for MLMs: while models approach human-level performance on vocabulary, they show clear deficits with possessives and even greater difficulty with demonstratives. Ablation analyses indicate that limitations in perspective-taking and spatial reasoning are key sources of these gaps. Instruction-based prompting reduces the gap for possessives but leaves demonstratives far below human performance. These results show that, unlike vocabulary, perspectival words pose a greater challenge in human communication, and this difficulty is amplified in MLMs, revealing a shortfall in their pragmatic and social-cognitive abilities.

Human preference data is essential for aligning large language models (LLMs) with human values, but collecting such data is often costly and inefficient-motivating the need for efficient data selection methods that reduce annotation costs while preserving alignment effectiveness. To address this issue, we propose Alignment Data Map, a data analysis tool for identifying and selecting effective preference data. We first evaluate alignment scores of the preference data by LLM-as-a-judge, explicit reward model, and reference-based approaches. The Alignment Data Map considers both response quality and inter-response variability based on the alignment scores. From our experimental findings, training on only 33% of samples that exhibit high-quality and low-variability, achieves comparable or superior alignment performance on MT-Bench, Evol-Instruct, and AlpacaEval, compared to training with the full dataset. In addition, Alignment Data Map detects potential label misannotations by analyzing correlations between annotated labels and alignment scores, improving annotation accuracy. The implementation is available at https://github.com/01choco/Alignment-Data-Map.

When a deep ReLU network is initialized with small weights, gradient descent (GD) is at first dominated by the saddle at the origin in parameter space. We study the so-called escape directions along which GD leaves the origin, which play a similar role as the eigenvectors of the Hessian for strict saddles. We show that the optimal escape direction features a low-rank bias in its deeper layers: the first singular value of the $\ell$-th layer weight matrix is at least $\ell^{\frac{1}{4}}$ larger than any other singular value. We also prove a number of related results about these escape directions. We suggest that deep ReLU networks exhibit saddle-to-saddle dynamics, with GD visiting a sequence of saddles with increasing bottleneck rank (Jacot, 2023).

A hallmark of human innovation is recombination -- the creation of novel ideas by integrating elements from existing concepts and mechanisms. In this work, we introduce CHIMERA, the first large-scale Knowledge Base (KB) of recombination examples automatically mined from the scientific literature. CHIMERA enables empirical analysis of how scientists recombine concepts and draw inspiration from different areas, and enables training models that propose cross-disciplinary research directions. To construct this KB, we define a new information extraction task: identifying recombination instances in papers. We curate an expert-annotated dataset and use it to fine-tune an LLM-based extraction model, which we apply to a broad corpus of AI papers. We also demonstrate generalization to a biological domain. We showcase the utility of CHIMERA through two applications. First, we analyze patterns of recombination across AI subfields. Second, we train a scientific hypothesis generation model using the KB, showing that it can propose directions that researchers rate as inspiring.

Large Language Models (LLMs) have extended their impact beyond Natural Language Processing, substantially fostering the development of interdisciplinary research. Recently, various LLM-based agents have been developed to assist scientific discovery progress across multiple aspects and domains. Among these, computer-using agents, capable of interacting with operating systems as humans do, are paving the way to automated scientific problem-solving and addressing routines in researchers' workflows. Recognizing the transformative potential of these agents, we introduce ScienceBoard, which encompasses two complementary contributions: (i) a realistic, multi-domain environment featuring dynamic and visually rich scientific workflows with integrated professional software, where agents can autonomously interact via different interfaces to accelerate complex research tasks and experiments; and (ii) a challenging benchmark of 169 high-quality, rigorously validated real-world tasks curated by humans, spanning scientific-discovery workflows in domains such as biochemistry, astronomy, and geoinformatics. Extensive evaluations of agents with state-of-the-art backbones (e.g., GPT-4o, Claude 3.7, UI-TARS) show that, despite some promising results, they still fall short of reliably assisting scientists in complex workflows, achieving only a 15% overall success rate. In-depth analysis further provides valuable insights for addressing current agent limitations and more effective design principles, paving the way to build more capable agents for scientific discovery. Our code, environment, and benchmark are at https://qiushisun.github.io/ScienceBoard-Home/.

Large Language Models (LLMs) have achieved remarkable performance across a wide range of mathematical benchmarks. However, concerns remain as to whether these successes reflect genuine reasoning or superficial pattern recognition. Existing evaluation methods, which typically focus either on the final answer or on the intermediate reasoning steps, reduce mathematical reasoning to a shallow input-output mapping, overlooking its inherently multi-stage and multi-dimensional cognitive nature. Inspired by Polya's problem-solving theory, we propose SMART, a benchmark that decomposes mathematical problem-solving into four cognitive dimensions: Semantic Understanding, Mathematical Reasoning, Arithmetic Computation, and Reflection & Refinement, and introduces dimension-specific tasks to measure the corresponding cognitive processes of LLMs. We apply SMART to 22 state-of-the-art open- and closed-source LLMs and uncover substantial discrepancies in their capabilities across dimensions. Our findings reveal genuine weaknesses in current models and motivate a new metric, the All-Pass Score, designed to better capture true problem-solving capability.

Log-likelihood vectors define a common space for comparing language models as probability distributions, enabling unified comparisons across heterogeneous settings. We extend this framework to training checkpoints and intermediate layers, and establish a consistent scale for KL divergence across pretraining, model size, random seeds, quantization, fine-tuning, and layers. Analysis of Pythia pretraining trajectories further shows that changes in log-likelihood space, as measured by the scaling behavior of KL divergence, are much smaller than in weight space, resulting in subdiffusive learning trajectories and early stabilization of language-model behavior despite weight drift.

Large language models (LLMs) are increasingly deployed for understanding large codebases, but whether they understand operational semantics of long code context or rely on pattern matching shortcuts remains unclear. We distinguish between lexical recall (retrieving code verbatim) and semantic recall (understanding operational semantics). Evaluating 10 state-of-the-art LLMs, we find that while frontier models achieve near-perfect, position-independent lexical recall, semantic recall degrades severely when code is centrally positioned in long contexts. We introduce semantic recall sensitivity to measure whether tasks require understanding of code's operational semantics vs. permit pattern matching shortcuts. Through a novel counterfactual measurement method, we show that models rely heavily on pattern matching shortcuts to solve existing code understanding benchmarks. We propose a new task SemTrace, which achieves high semantic recall sensitivity through unpredictable operations; LLMs' accuracy exhibits severe positional effects, with median accuracy drops of 92.73% versus CRUXEval's 53.36% as the relevant code snippet approaches the middle of the input code context. Our findings suggest current evaluations substantially underestimate semantic recall failures in long context code understanding.

The assessment of evaluation metrics (meta-evaluation) is crucial for determining the suitability of existing metrics in text-to-image (T2I) generation tasks. Human-based meta-evaluation is costly and time-intensive, and automated alternatives are scarce. We address this gap and propose CROC: a scalable framework for automated Contrastive Robustness Checks that systematically probes and quantifies metric robustness by synthesizing contrastive test cases across a comprehensive taxonomy of image properties. With CROC, we generate a pseudo-labeled dataset (CROC$^{syn}$) of over 1 million contrastive prompt-image pairs to enable a fine-grained comparison of evaluation metrics. We also use this dataset to train CROCScore, a new metric that achieves state-of-the-art performance among open-source methods, demonstrating an additional key application of our framework. To complement this dataset, we introduce a human-supervised benchmark (CROC$^{hum}$) targeting especially challenging categories. Our results highlight robustness issues in existing metrics: for example, many fail on prompts involving negation, and all tested open-source metrics fail on at least 24% of cases involving correct identification of body parts.