Front-end engineering involves a complex workflow where engineers conceptualize designs, translate them into code, and iteratively refine the implementation. While recent benchmarks primarily focus on converting visual designs to code, we present FullFront, a benchmark designed to evaluate Multimodal Large Language Models (MLLMs) \textbf{across the full front-end development pipeline}. FullFront assesses three fundamental tasks that map directly to the front-end engineering pipeline: Webpage Design (conceptualization phase), Webpage Perception QA (comprehension of visual organization and elements), and Webpage Code Generation (implementation phase). Unlike existing benchmarks that use either scraped websites with bloated code or oversimplified LLM-generated HTML, FullFront employs a novel, two-stage process to transform real-world webpages into clean, standardized HTML while maintaining diverse visual designs and avoiding copyright issues. Extensive testing of state-of-the-art MLLMs reveals significant limitations in page perception, code generation (particularly for image handling and layout), and interaction implementation. Our results quantitatively demonstrate performance disparities across models and tasks, and highlight a substantial gap between current MLLM capabilities and human expert performance in front-end engineering. The FullFront benchmark and code are available in https://github.com/Mikivishy/FullFront.

Large language models (LLMs) have demonstrated strong performance in general-purpose machine translation, but their effectiveness in complex, domain-sensitive translation tasks remains underexplored. Recent advancements in Large Reasoning Models (LRMs), raise the question of whether structured reasoning can enhance translation quality across diverse domains. In this work, we compare the performance of LRMs with traditional LLMs across 15 representative domains and four translation directions. Our evaluation considers various factors, including task difficulty, input length, and terminology density. We use a combination of automatic metrics and an enhanced MQM-based evaluation hierarchy to assess translation quality. Our findings show that LRMs consistently outperform traditional LLMs in semantically complex domains, especially in long-text and high-difficulty translation scenarios. Moreover, domain-adaptive prompting strategies further improve performance by better leveraging the reasoning capabilities of LRMs. These results highlight the potential of structured reasoning in MDMT tasks and provide valuable insights for optimizing translation systems in domain-sensitive contexts.

Human-annotated data plays a vital role in training large language models (LLMs), such as supervised fine-tuning and human preference alignment. However, it is not guaranteed that paid human annotators produce high-quality data. In this paper, we study how to incentivize human annotators to do so. We start from a principal-agent model to model the dynamics between the company (the principal) and the annotator (the agent), where the principal can only monitor the annotation quality by examining $n$ samples. We investigate the maximum likelihood estimators (MLE) and the corresponding hypothesis testing to incentivize annotators: the agent is given a bonus if the MLE passes the test. By analyzing the variance of the outcome, we show that the strategic behavior of the agent makes the hypothesis testing very different from traditional ones: Unlike the exponential rate proved by the large deviation theory, the principal-agent model's hypothesis testing rate is of $Θ(1/\sqrt{n \log n})$. Our theory implies two criteria for the \emph{golden questions} to monitor the performance of the annotators: they should be of (1) high certainty and (2) similar format to normal ones. In that light, we select a set of golden questions in human preference data. By doing incentive-compatible experiments, we find out that the annotators' behavior is better revealed by those golden questions, compared to traditional survey techniques such as instructed manipulation checks.

Selecting artificial intelligence (AI) models, such as large language models (LLMs), from multiple candidates requires accurate performance estimation. This is ideally achieved through empirical evaluations involving abundant real-world data. However, such evaluations are costly and impractical at scale. To address this challenge, autoevaluation methods leverage synthetic data produced by automated evaluators, such as LLMs-as-judges, reducing variance but potentially introducing bias. Recent approaches have employed semi-supervised prediction-powered inference (\texttt{PPI}) to correct for the bias of autoevaluators. However, the use of autoevaluators may lead in practice to a degradation in sample efficiency compared to conventional methods using only real-world data. In this paper, we propose \texttt{R-AutoEval+}, a novel framework that provides finite-sample reliability guarantees on the model evaluation, while also ensuring an enhanced (or at least no worse) sample efficiency compared to conventional methods. The key innovation of \texttt{R-AutoEval+} is an adaptive construction of the model evaluation variable, which dynamically tunes its reliance on synthetic data, reverting to conventional methods when the autoevaluator is insufficiently accurate. Experiments on the use of LLMs-as-judges for the optimization of quantization settings for the weights of an LLM, and for prompt design in LLMs confirm the reliability and efficiency of \texttt{R-AutoEval+}.

Sparse Autoencoders (SAEs) are a prominent tool in mechanistic interpretability (MI) for decomposing neural network activations into interpretable features. However, the aspiration to identify a canonical set of features is challenged by the observed inconsistency of learned SAE features across different training runs, undermining the reliability and efficiency of MI research. This position paper argues that mechanistic interpretability should prioritize feature consistency in SAEs -- the reliable convergence to equivalent feature sets across independent runs. We propose using the Pairwise Dictionary Mean Correlation Coefficient (PW-MCC) as a practical metric to operationalize consistency and demonstrate that high levels are achievable (0.80 for TopK SAEs on LLM activations) with appropriate architectural choices. Our contributions include detailing the benefits of prioritizing consistency; providing theoretical grounding and synthetic validation using a model organism, which verifies PW-MCC as a reliable proxy for ground-truth recovery; and extending these findings to real-world LLM data, where high feature consistency strongly correlates with the semantic similarity of learned feature explanations. We call for a community-wide shift towards systematically measuring feature consistency to foster robust cumulative progress in MI.

Representation intervention aims to locate and modify the representations that encode the underlying concepts in Large Language Models (LLMs) to elicit the aligned and expected behaviors. Despite the empirical success, it has never been examined whether one could locate the faithful concepts for intervention. In this work, we explore the question in safety alignment. If the interventions are faithful, the intervened LLMs should erase the harmful concepts and be robust to both in-distribution adversarial prompts and the out-of-distribution (OOD) jailbreaks. While it is feasible to erase harmful concepts without degrading the benign functionalities of LLMs in linear settings, we show that it is infeasible in the general non-linear setting. To tackle the issue, we propose Concept Concentration (COCA). Instead of identifying the faithful locations to intervene, COCA refractors the training data with an explicit reasoning process, which firstly identifies the potential unsafe concepts and then decides the responses. Essentially, COCA simplifies the decision boundary between harmful and benign representations, enabling more effective linear erasure. Extensive experiments with multiple representation intervention methods and model architectures demonstrate that COCA significantly reduces both in-distribution and OOD jailbreak success rates, and meanwhile maintaining strong performance on regular tasks such as math and code generation.

Accelerator memory and networking constraints have emerged as dominant bottlenecks when training large language models LLMs with billions of parameters. Existing low rank gradient estimators such as GaLoRE and FLORA compress gradients and optimizer tensors by projecting weight gradients onto a rank r subspace, enabling LLM training on consumer hardware. Yet, these methods are either biased or subject to high estimator variance. Moreover, the optimizer state based on the first and second moments estimates expressed in the previous subspace becomes misaligned whenever the projection is updated, leading to instabilities during training. We propose PLUMAGE: Probabilistic Low rank Unbiased Minimum vAriance Gradient Estimator. PLUMAGE is a drop in replacement for existing low rank gradient estimators. It does not introduce new hyperparameters beyond the chosen rank r and the update interval. In addition, we resolve optimizer state misalignment issues to prevent spurious weight updates and enhance training stability. We empirically demonstrate that PLUMAGE shrinks the full rank optimization's gap over the pre training evaluation loss by 33% on average across models and the average training loss across the GLUE benchmark by 28% within a similar computational and memory footprint as GaloRE.

Despite LLMs' explicit alignment against demographic stereotypes, they have been shown to exhibit biases under various social contexts. In this work, we find that LLMs exhibit concerning biases in how they associate solution veracity with demographics. Through experiments across five human value-aligned LLMs on mathematics, coding, commonsense, and writing problems, we reveal two forms of such veracity biases: Attribution Bias, where models disproportionately attribute correct solutions to certain demographic groups, and Evaluation Bias, where models' assessment of identical solutions varies based on perceived demographic authorship. Our results show pervasive biases: LLMs consistently attribute fewer correct solutions and more incorrect ones to African-American groups in math and coding, while Asian authorships are least preferred in writing evaluation. In additional studies, we show LLMs automatically assign racially stereotypical colors to demographic groups in visualization code, suggesting these biases are deeply embedded in models' reasoning processes. Our findings indicate that demographic bias extends beyond surface-level stereotypes and social context provocations, raising concerns about LLMs' deployment in educational and evaluation settings.

Large Language Models (LLMs) have demonstrated effectiveness in code generation tasks. To enable LLMs to address more complex coding challenges, existing research has focused on crafting multi-agent systems with agentic workflows, where complex coding tasks are decomposed into sub-tasks, assigned to specialized agents. Despite their effectiveness, current approaches heavily rely on hand-crafted agentic workflows, with both agent topologies and prompts manually designed, which limits their ability to automatically adapt to different types of coding problems. To address these limitations and enable automated workflow design, we propose \textbf{S}elf-\textbf{E}volving \textbf{W}orkflow (\textbf{SEW}), a novel self-evolving framework that automatically generates and optimises multi-agent workflows. Extensive experiments on three coding benchmark datasets, including the challenging LiveCodeBench, demonstrate that our SEW can automatically design agentic workflows and optimise them through self-evolution, bringing up to 33\% improvement on LiveCodeBench compared to using the backbone LLM only. Furthermore, by investigating different representation schemes of workflow, we provide insights into the optimal way to encode workflow information with text.

Large language models (LLMs) have shown remarkable generalization across tasks, leading to increased interest in integrating speech with LLMs. These speech LLMs (SLLMs) typically use supervised fine-tuning to align speech with text-based LLMs. However, the lack of annotated speech data across a wide range of tasks hinders alignment efficiency, resulting in poor generalization. To address these issues, we propose a novel multi-task 'behavior imitation' method with speech-text interleaving, called MTBI, which relies solely on paired speech and transcripts. By ensuring the LLM decoder generates equivalent responses to paired speech and text, we achieve a more generalized SLLM. Interleaving is used to further enhance alignment efficiency. We introduce a simple benchmark to evaluate prompt and task generalization across different models. Experimental results demonstrate that our MTBI outperforms SOTA SLLMs on both prompt and task generalization, while requiring less supervised speech data.