Simulation-based testing has become a crucial complement to road testing for ensuring the safety of cyber physical systems (CPS). As a result, significant research efforts have been directed toward identifying failure scenarios within simulation environments. However, a critical question remains. Are the AV failure scenarios discovered in simulation reproducible on actual systems in the real world? The sim-to-real gap caused by differences between simulated and real sensor data means that failure scenarios identified in simulation might either be artifacts of synthetic sensor data or actual issues that also occur with real sensor data. To address this, an effective approach to validating simulated failure scenarios is to locate occurrences of these scenarios within real-world datasets and verify whether the failure persists on the datasets. To this end, we introduce a formal definition of how labeled time series sensor data can match an abstract scenario, represented as a scenario program using the Scenic probabilistic programming language. We present a querying algorithm that, given a scenario program and a labeled dataset, identifies the subset of data that matches the specified scenario. Our experiment shows that our algorithm is more accurate and orders of magnitude faster in querying scenarios than the state-of-the-art commercial vision large language models, and can scale with the duration of queried time series data.

Large Language Models (LLMs) have demonstrated remarkable reasoning abilities, yet existing test-time frameworks often rely on coarse self-verification and self-correction, limiting their effectiveness on complex tasks. In this paper, we propose Socratic Self-Refine (SSR), a novel framework for fine-grained evaluation and precise refinement of LLM reasoning. Our proposed SSR decomposes model responses into verifiable (sub-question, sub-answer) pairs, enabling step-level confidence estimation through controlled re-solving and self-consistency checks. By pinpointing unreliable steps and iteratively refining them, SSR produces more accurate and interpretable reasoning chains. Empirical results across five reasoning benchmarks and three LLMs show that SSR consistently outperforms state-of-the-art iterative self-refinement baselines. Beyond performance gains, SSR provides a principled black-box approach for evaluating and understanding the internal reasoning processes of LLMs. Code is available at https://github.com/SalesforceAIResearch/socratic-self-refine-reasoning.

Language prediction is constrained by informational entropy intrinsic to language, such that there exists a limit to how accurate any language model can become and equivalently a lower bound to language compression. The most efficient language compression algorithms today are causal (next token prediction) large language models, but the use of these models to form accurate estimates of language entropy is currently computationally infeasible. We introduce encoder-augmented causal decoder model architectures that exhibit superior training efficiency characteristics and achieve higher compression than causal transformers even when trained on modest hardware. We demonstrate how entropy estimates can be obtained on a per-token basis, and show that the generalization of models trained to approach the entropy of their training data necessarily exceeds the generalization of models trained to minimize loss beyond this value. We show empirically that causal models trained to approach but not exceed estimated per-token entropies exhibit greater generalization than models trained without taking entropy into account.

Large Vision-Language Models (VLMs) excel at understanding and generating video descriptions but their high memory, computation, and deployment demands hinder practical use particularly for blind and low-vision (BLV) users who depend on detailed, context-aware descriptions. To study the effect of model size on accessibility-focused description quality, we evaluate SmolVLM2 variants with 500M and 2.2B parameters across two diverse datasets: AVCaps (outdoor), and Charades (indoor). In this work, we introduce two novel evaluation frameworks specifically designed for BLV accessibility assessment: the Multi-Context BLV Framework evaluating spatial orientation, social interaction, action events, and ambience contexts; and the Navigational Assistance Framework focusing on mobility-critical information. Additionally, we conduct a systematic evaluation of four different prompt design strategies and deploy both models on a smartphone, evaluating FP32 and INT8 precision variants to assess real-world performance constraints on resource-limited mobile devices.

The WikiRace game, where players navigate between Wikipedia articles using only hyperlinks, serves as a compelling benchmark for goal-directed search in complex information networks. This paper presents a systematic evaluation of navigation strategies for this task, comparing agents guided by graph-theoretic structure (betweenness centrality), semantic meaning (language model embeddings), and hybrid approaches. Through rigorous benchmarking on a large Wikipedia sub-graph, we demonstrate that a purely greedy agent guided by the semantic similarity of article titles is overwhelmingly effective. This strategy, when combined with a simple loop-avoidance mechanism, achieved a perfect success rate and navigated the network with an efficiency an order of magnitude better than structural or hybrid methods. Our findings highlight the critical limitations of purely structural heuristics for goal-directed search and underscore the transformative potential of large language models to act as powerful, zero-shot semantic navigators in complex information spaces.

Recent multimodal large language models (MLLMs) still struggle with long document understanding due to two fundamental challenges: information interference from abundant irrelevant content, and the quadratic computational cost of Transformer-based architectures. Existing approaches primarily fall into two categories: token compression, which sacrifices fine-grained details; and introducing external retrievers, which increase system complexity and prevent end-to-end optimization. To address these issues, we conduct an in-depth analysis and observe that MLLMs exhibit a human-like coarse-to-fine reasoning pattern: early Transformer layers attend broadly across the document, while deeper layers focus on relevant evidence pages. Motivated by this insight, we posit that the inherent evidence localization capabilities of MLLMs can be explicitly leveraged to perform retrieval during the reasoning process, facilitating efficient long document understanding. To this end, we propose URaG, a simple-yet-effective framework that Unifies Retrieval and Generation within a single MLLM. URaG introduces a lightweight cross-modal retrieval module that converts the early Transformer layers into an efficient evidence selector, identifying and preserving the most relevant pages while discarding irrelevant content. This design enables the deeper layers to concentrate computational resources on pertinent information, improving both accuracy and efficiency. Extensive experiments demonstrate that URaG achieves state-of-the-art performance while reducing computational overhead by 44-56%. The code is available at https://github.com/shi-yx/URaG.

We introduce a comprehensive benchmark for conversational memory evaluation containing 75,336 question-answer pairs across diverse categories including user facts, assistant recall, abstention, preferences, temporal changes, and implicit connections. While existing benchmarks have advanced the field, our work addresses fundamental challenges in statistical power, data generation consistency, and evaluation flexibility that limit current memory evaluation frameworks. We examine the relationship between conversational memory and retrieval-augmented generation (RAG). While these systems share fundamental architectural patterns--temporal reasoning, implicit extraction, knowledge updates, and graph representations--memory systems have a unique characteristic: they start from zero and grow progressively with each conversation. This characteristic enables naive approaches that would be impractical for traditional RAG. Consistent with recent findings on long context effectiveness, we observe that simple full-context approaches achieve 70-82% accuracy even on our most challenging multi-message evidence cases, while sophisticated RAG-based memory systems like Mem0 achieve only 30-45% when operating on conversation histories under 150 interactions. Our analysis reveals practical transition points: long context excels for the first 30 conversations, remains viable with manageable trade-offs up to 150 conversations, and typically requires hybrid or RAG approaches beyond that point as costs and latencies become prohibitive. These patterns indicate that the small-corpus advantage of conversational memory--where exhaustive search and complete reranking are feasible--deserves dedicated research attention rather than simply applying general RAG solutions to conversation histories.

Large Language Models (LLMs) are commonly evaluated for robustness against paraphrased or semantically equivalent jailbreak prompts, yet little attention has been paid to linguistic variation as an attack surface. In this work, we systematically study how linguistic styles such as fear or curiosity can reframe harmful intent and elicit unsafe responses from aligned models. We construct style-augmented jailbreak benchmark by transforming prompts from 3 standard datasets into 11 distinct linguistic styles using handcrafted templates and LLM-based rewrites, while preserving semantic intent. Evaluating 16 open- and close-source instruction-tuned models, we find that stylistic reframing increases jailbreak success rates by up to +57 percentage points. Styles such as fearful, curious and compassionate are most effective and contextualized rewrites outperform templated variants. To mitigate this, we introduce a style neutralization preprocessing step using a secondary LLM to strip manipulative stylistic cues from user inputs, significantly reducing jailbreak success rates. Our findings reveal a systemic and scaling-resistant vulnerability overlooked in current safety pipelines.

Olympiad-level physics problem-solving presents a significant challenge for both humans and artificial intelligence (AI), as it requires a sophisticated integration of precise calculation, abstract reasoning, and a fundamental grasp of physical principles. The Chinese Physics Olympiad (CPhO), renowned for its complexity and depth, serves as an ideal and rigorous testbed for these advanced capabilities. In this paper, we introduce LOCA-R (LOgical Chain Augmentation for Reasoning), an improved version of the LOCA framework adapted for complex reasoning, and apply it to the CPhO 2025 theory examination. LOCA-R achieves a near-perfect score of 313 out of 320 points, solidly surpassing the highest-scoring human competitor and significantly outperforming all baseline methods. Physics problem-solving, or more generally complex scientific reasoning, stands as a challenging frontier for both humans and AI. It demands not only mathematical derivations, but also the ability to translate complex real-world scenarios described in natural language into abstract models. This process requires a deep understanding of physical laws to select and apply the appropriate principles. Therefore, given the strict demand for accuracy in scientific problems, the most critical task is to maximize the problem-solving capabilities of LLMs, pushing them towards perfect scores. The trade-off between performance and computational cost, while important, is a subsequent consideration.

Abstract--While prompt optimization has emerged as a critical technique for enhancing language model performance, existing approaches primarily focus on elicitation-based strategies that search for optimal prompts to activate models' capabilities. These methods exhibit fundamental limitations when addressing knowledge-intensive tasks, as they operate within fixed parametric boundaries rather than providing the factual knowledge, terminology precision, and reasoning patterns required in specialized domains. T o address these limitations, we propose Knowledge-Provision-based Prompt Optimization (KPPO), a framework that reformulates prompt optimization as systematic knowledge integration rather than potential elicitation. KPPO introduces three key innovations: 1) a knowledge gap filling mechanism for knowledge gap identification and targeted remediation; 2) a batch-wise candidate evaluation approach that considers both performance improvement and distributional stability; 3) an adaptive knowledge pruning strategy that balances performance and token efficiency, reducing up to 29% token usage. Extensive evaluation on 15 knowledge-intensive benchmarks from various domains demonstrates KPPO's superiority over elicitation-based methods, with an average performance improvement of ~6% over the strongest baseline while achieving comparable or lower token consumption. Large Language Models (LLMs) have achieved unprecedented performance across diverse natural language processing tasks through sophisticated prompt engineering techniques [1]. The field has evolved from manual prompt design approaches [2], [3] to automated optimization frameworks [4]- [7], where optimizer LLMs iteratively refine prompts to maximize task performance. These automated approaches, collectively termed elicitation-based optimization, operate under the fundamental assumption that optimal prompts can unlock latent capabilities within pre-trained model parameters through strategic reformulation of instructions, exemplars, or reasoning templates.