Neural Networks (NNs) trained through supervised learning struggle with managing edge-case scenarios common in real-world driving due to the intractability of exhaustive datasets covering all edge-cases, making knowledge-driven approaches, akin to how humans intuitively detect unexpected driving behavior, a suitable complement to data-driven methods. This work proposes a hybrid architecture combining low-level Model Predictive Controller (MPC) with locally deployed Large Language Models (LLMs) to enhance decision-making and Human Machine Interaction (HMI). The DecisionxLLM module evaluates robotic state information against natural language instructions to ensure adherence to desired driving behavior. The MPCxLLM module then adjusts MPC parameters based on LLM-generated insights, achieving control adaptability while preserving the safety and constraint guarantees of traditional MPC systems. Further, to enable efficient on-board deployment and to eliminate dependency on cloud connectivity, we shift processing to the on-board computing platform: We propose an approach that exploits Retrieval Augmented Generation (RAG), Low Rank Adaptation (LoRA) fine-tuning, and quantization. Experimental results demonstrate that these enhancements yield significant improvements in reasoning accuracy by up to 10.45%, control adaptability by as much as 52.2%, and up to 10.5x increase in computational efficiency (tokens/s), validating the proposed framework's practicality for real-time deployment even on down-scaled robotic platforms. This work bridges high-level decision-making with low-level control adaptability, offering a synergistic framework for knowledge-driven and adaptive Autonomous Driving Systems (ADS).

Timing analysis is an essential and demanding verification method for Very Large Scale Integrated (VLSI) circuit design and optimization. In addition, it also serves as the cornerstone of the final sign-off, determining whether the chip is ready to be sent to the semiconductor foundry for fabrication. Recently, as the technology advance relentlessly, smaller metal pitches and the increasing number of devices have led to greater challenges and longer turn-around-time for experienced human designers to debug timing issues from the Multi-Corner Multi-Mode (MCMM) timing reports. As a result, an efficient and intelligent methodology is highly necessary and essential for debugging timing issues and reduce the turnaround times. Recently, Large Language Models (LLMs) have shown great promise across various tasks in language understanding and interactive decision-making, incorporating reasoning and actions. In this work, we propose a timing analysis agent, that is empowered by multi-LLMs task solving, and incorporates a novel hierarchical planning and solving flow to automate the analysis of timing reports from commercial tool. In addition, we build a Timing Debug Relation Graph (TDRG) that connects the reports with the relationships of debug traces from experienced timing engineers. The timing analysis agent employs the novel Agentic Retrieval Augmented Generation (RAG) approach, that includes agent and coding to retrieve data accurately, on the developed TDRG. In our studies, the proposed timing analysis agent achieves an average 98% pass-rate on a single-report benchmark and a 90% pass-rate for multi-report benchmark from industrial designs, demonstrating its effectiveness and adaptability.

This paper presents a proposal for the governance of frontier AI systems through a hybrid public-private system. Private bodies, authorized and overseen by government, provide certifications to developers of frontier AI systems on an opt-in basis. In exchange for opting in, frontier AI firms receive protections from tort liability for customer misuse of their models. Before detailing the proposal, the paper explores more commonly discussed approaches to AI governance, analyzing their strengths and flaws. It also examines the nature of frontier AI governance itself. The paper includes consideration of the political economic, institutional, legal, safety, and other merits and tradeoffs inherent in the governance system it proposes.

Existing computer vision(CV)-based structural damage identification models demonstrate notable accuracy in categorizing and localizing damage. However, these models present several critical limitations that hinder their practical application in civil engineering(CE). Primarily, their ability to recognize damage types remains constrained, preventing comprehensive analysis of the highly varied and complex conditions encountered in real-world CE structures. Second, these models lack linguistic capabilities, rendering them unable to articulate structural damage characteristics through natural language descriptions. With the continuous advancement of artificial intelligence(AI), large multi-modal models(LMMs) have emerged as a transformative solution, enabling the unified encoding and alignment of textual and visual data. These models can autonomously generate detailed descriptive narratives of structural damage while demonstrating robust generalization across diverse scenarios and tasks. This study introduces SDIGLM, an innovative LMM for structural damage identification, developed based on the open-source VisualGLM-6B architecture. To address the challenge of adapting LMMs to the intricate and varied operating conditions in CE, this work integrates a U-Net-based semantic segmentation module to generate defect segmentation maps as visual Chain of Thought(CoT). Additionally, a multi-round dialogue fine-tuning dataset is constructed to enhance logical reasoning, complemented by a language CoT formed through prompt engineering. By leveraging this multi-modal CoT, SDIGLM surpasses general-purpose LMMs in structural damage identification, achieving an accuracy of 95.24% across various infrastructure types. Moreover, the model effectively describes damage characteristics such as hole size, crack direction, and corrosion severity.

This work revisits the dominant supervised fine-tuning (SFT) then reinforcement learning (RL) paradigm for training Large Vision-Language Models (LVLMs), and reveals a key finding: SFT can significantly undermine subsequent RL by inducing ``pseudo reasoning paths'' imitated from expert models. While these paths may resemble the native reasoning paths of RL models, they often involve prolonged, hesitant, less informative steps, and incorrect reasoning. To systematically study this effect, we introduce VLAA-Thinking, a new multimodal dataset designed to support reasoning in LVLMs. Constructed via a six-step pipeline involving captioning, reasoning distillation, answer rewrite and verification, VLAA-Thinking comprises high-quality, step-by-step visual reasoning traces for SFT, along with a more challenging RL split from the same data source. Using this dataset, we conduct extensive experiments comparing SFT, RL and their combinations. Results show that while SFT helps models learn reasoning formats, it often locks aligned models into imitative, rigid reasoning modes that impede further learning. In contrast, building on the Group Relative Policy Optimization (GRPO) with a novel mixed reward module integrating both perception and cognition signals, our RL approach fosters more genuine, adaptive reasoning behavior. Notably, our model VLAA-Thinker, based on Qwen2.5VL 3B, achieves top-1 performance on Open LMM Reasoning Leaderboard (https://huggingface.co/spaces/opencompass/Open_LMM_Reasoning_Leaderboard) among 4B scale LVLMs, surpassing the previous state-of-the-art by 1.8%. We hope our findings provide valuable insights in developing reasoning-capable LVLMs and can inform future research in this area.

Visual Language Models (VLMs) show remarkable performance in visual reasoning tasks, successfully tackling college-level challenges that require high-level understanding of images. However, some recent reports of VLMs struggling to reason about elemental visual concepts like orientation, position, continuity, and occlusion suggest a potential gulf between human and VLM vision. Here we use the toolkit of neuropsychology to systematically assess the capabilities of three state-of-the-art VLMs across visual domains. Using 51 tests drawn from six clinical and experimental batteries, we characterise the visual abilities of leading VLMs relative to normative performance in healthy adults. While the models excel in straightforward object recognition tasks, we find widespread deficits in low- and mid-level visual abilities that would be considered clinically significant in humans. These selective deficits, profiled through validated test batteries, suggest that an artificial system can achieve complex object recognition without developing foundational visual concepts that in humans require no explicit training.

An AI system can create and maintain knowledge only to the extent that it can verify that knowledge itself. Recent work on long Chain-of-Thought reasoning has demonstrated great potential of LLMs on solving competitive problems, but their verification ability remains to be weak and not sufficiently investigated. In this paper, we propose Heimdall, the long CoT verification LLM that can accurately judge the correctness of solutions. With pure reinforcement learning, we boost the verification accuracy from 62.5% to 94.5% on competitive math problems. By scaling with repeated sampling, the accuracy further increases to 97.5%. Through human evaluation, Heimdall demonstrates impressive generalization capabilities, successfully detecting most issues in challenging math proofs, the type of which is not included during training. Furthermore, we propose Pessimistic Verification to extend the functionality of Heimdall to scaling up the problem solving. It calls Heimdall to judge the solutions from a solver model and based on the pessimistic principle, selects the most likely correct solution with the least uncertainty. Taking DeepSeek-R1-Distill-Qwen-32B as the solver model, Pessimistic Verification improves the solution accuracy on AIME2025 from 54.2% to 70.0% with 16x compute budget and to 83.3% with more compute budget. With the stronger solver Gemini 2.5 Pro, the score reaches 93.0%. Finally, we prototype an automatic knowledge discovery system, a ternary system where one poses questions, another provides solutions, and the third verifies the solutions. Using the data synthesis work NuminaMath for the first two components, Heimdall effectively identifies problematic records within the dataset and reveals that nearly half of the data is flawed, which interestingly aligns with the recent ablation studies from NuminaMath.

Large language model unlearning has become a critical challenge in ensuring safety and controlled model behavior by removing undesired data-model influences from the pretrained model while preserving general utility. Significant recent efforts have been dedicated to developing LLM unlearning benchmarks such as WMDP (Weapons of Mass Destruction Proxy) and MUSE (Machine Unlearning Six-way Evaluation), facilitating standardized unlearning performance assessment and method comparison. Despite their usefulness, we uncover for the first time a novel coreset effect within these benchmarks. Specifically, we find that LLM unlearning achieved with the original (full) forget set can be effectively maintained using a significantly smaller subset (functioning as a "coreset"), e.g., as little as 5% of the forget set, even when selected at random. This suggests that LLM unlearning in these benchmarks can be performed surprisingly easily, even in an extremely low-data regime. We demonstrate that this coreset effect remains strong, regardless of the LLM unlearning method used, such as NPO (Negative Preference Optimization) and RMU (Representation Misdirection Unlearning), the popular ones in these benchmarks. The surprisingly strong coreset effect is also robust across various data selection methods, ranging from random selection to more sophisticated heuristic approaches. We explain the coreset effect in LLM unlearning through a keyword-based perspective, showing that keywords extracted from the forget set alone contribute significantly to unlearning effectiveness and indicating that current unlearning is driven by a compact set of high-impact tokens rather than the entire dataset. We further justify the faithfulness of coreset-unlearned models along additional dimensions, such as mode connectivity and robustness to jailbreaking attacks. Codes are available at https://github.com/OPTML-Group/MU-Coreset.

Despite the successes of language models, their evaluation remains a daunting challenge for new and existing tasks. We consider the task of text simplification, commonly used to improve information accessibility, where evaluation faces two major challenges. First, the data in existing benchmarks might not reflect the capabilities of current language models on the task, often containing disfluent, incoherent, or simplistic examples. Second, existing human ratings associated with the benchmarks often contain a high degree of disagreement, resulting in inconsistent ratings; nevertheless, existing metrics still have to show higher correlations with these imperfect ratings. As a result, evaluation for the task is not reliable and does not reflect expected trends (e.g., more powerful models being assigned higher scores). We address these challenges for the task of text simplification through three contributions. First, we introduce SynthSimpliEval, a synthetic benchmark for text simplification featuring simplified sentences generated by models of varying sizes. Through a pilot study, we show that human ratings on our benchmark exhibit high inter-annotator agreement and reflect the expected trend: larger models produce higher-quality simplifications. Second, we show that auto-evaluation with a panel of LLM judges (LLMs-as-a-jury) often suffices to obtain consistent ratings for the evaluation of text simplification. Third, we demonstrate that existing learnable metrics for text simplification benefit from training on our LLMs-as-a-jury-rated synthetic data, closing the gap with pure LLMs-as-a-jury for evaluation. Overall, through our case study on text simplification, we show that a reliable evaluation requires higher quality test data, which could be obtained through synthetic data and LLMs-as-a-jury ratings.

Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks, with their performance heavily dependent on the quality of input prompts. While prompt engineering has proven effective, it typically relies on manual adjustments, making it time-consuming and potentially suboptimal. This paper introduces GAAPO (Genetic Algorithm Applied to Prompt Optimization), a novel hybrid optimization framework that leverages genetic algorithm principles to evolve prompts through successive generations. Unlike traditional genetic approaches that rely solely on mutation and crossover operations, GAAPO integrates multiple specialized prompt generation strategies within its evolutionary framework. Through extensive experimentation on diverse datasets including ETHOS, MMLU-Pro, and GPQA, our analysis reveals several important point for the future development of automatic prompt optimization methods: importance of the tradeoff between the population size and the number of generations, effect of selection methods on stability results, capacity of different LLMs and especially reasoning models to be able to automatically generate prompts from similar queries... Furthermore, we provide insights into the relative effectiveness of different prompt generation strategies and their evolution across optimization phases. These findings contribute to both the theoretical understanding of prompt optimization and practical applications in improving LLM performance.