Harnessing large offline datasets is vital for training foundation models that can generalize across diverse tasks. Offline Reinforcement Learning (RL) offers a powerful framework for these scenarios, enabling the derivation of optimal policies even from suboptimal data. The Prompting Decision Transformer (PDT) is an offline RL multi-task model that distinguishes tasks through stochastic trajectory prompts, which are task-specific tokens maintained in context during rollouts. However, PDT samples these tokens uniformly at random from per-task demonstration datasets, failing to account for differences in token informativeness and potentially leading to performance degradation. To address this limitation, we introduce a scalable bandit-based prompt-tuning method that dynamically learns to construct high-performance trajectory prompts. Our approach significantly enhances downstream task performance without modifying the pre-trained Transformer backbone. Empirical results on benchmark tasks and a newly designed multi-task environment demonstrate the effectiveness of our method, creating a seamless bridge between general multi-task offline pre-training and task-specific online adaptation.

Entity retrieval plays a crucial role in the utilization of Electronic Health Records (EHRs) and is applied across a wide range of clinical practices. However, a comprehensive evaluation of this task is lacking due to the absence of a public benchmark. In this paper, we propose the development and release of a novel benchmark for evaluating entity retrieval in EHRs, with a particular focus on the semantic gap issue. Using discharge summaries from the MIMIC-III dataset, we incorporate ICD codes and prescription labels associated with the notes as queries, and annotate relevance judgments using GPT-4. In total, we use 1,000 patient notes, generate 1,246 queries, and provide over 77,000 relevance annotations. To offer the first assessment of the semantic gap, we introduce a novel classification system for relevance matches. Leveraging GPT-4, we categorize each relevant pair into one of five categories: string, synonym, abbreviation, hyponym, and implication. Using the proposed benchmark, we evaluate several retrieval methods, including BM25, query expansion, and state-of-the-art dense retrievers. Our findings show that BM25 provides a strong baseline but struggles with semantic matches. Query expansion significantly improves performance, though it slightly reduces string match capabilities. Dense retrievers outperform traditional methods, particularly for semantic matches, and general-domain dense retrievers often surpass those trained specifically in the biomedical domain.

A fundamental issue in deep learning has been adversarial robustness. As these systems have scaled, such issues have persisted. Currently, large language models (LLMs) with billions of parameters suffer from adversarial attacks just like their earlier, smaller counterparts. However, the threat models have changed. Previously, having gray-box access, where input embeddings or output logits/probabilities were visible to the user, might have been reasonable. However, with the introduction of closed-source models, no information about the model is available apart from the generated output. This means that current black-box attacks can only utilize the final prediction to detect if an attack is successful. In this work, we investigate and demonstrate the potential of attack guidance, akin to using output probabilities, while having only black-box access in a classification setting. This is achieved through the ability to elicit confidence from the model. We empirically show that the elicited confidence is calibrated and not hallucinated for current LLMs. By minimizing the elicited confidence, we can therefore increase the likelihood of misclassification. Our new proposed paradigm demonstrates promising state-of-the-art results on three datasets across two models (LLaMA-3-8B-Instruct and Mistral-7B-Instruct-V0.3) when comparing our technique to existing hard-label black-box attack methods that introduce word-level substitutions.

Advancing machine visual reasoning requires a deeper understanding of how Vision-Language Models (VLMs) process and interpret complex visual patterns. This work introduces a novel, cognitively-inspired evaluation framework to systematically analyze VLM reasoning on natural image-based Bongard Problems. We propose three structured paradigms -- Direct Visual Rule Learning, Deductive Rule Learning, and Componential Analysis -- designed to progressively enforce step-wise reasoning and disentangle the interplay between perception and reasoning. Our evaluation shows that advanced, closed-source VLMs (GPT-4o and Gemini 2.0) achieve near-superhuman performance, particularly when provided with high-quality image descriptions, while open-source models exhibit a significant performance bottleneck due to deficiencies in perception. An ablation study further confirms that perception, rather than reasoning, is the primary limiting factor, as open-source models apply extracted rules effectively when given accurate descriptions. These findings underscore the critical role of robust multimodal perception in enhancing generalizable visual reasoning and highlight the importance of structured, step-wise reasoning paradigms for advancing machine intelligence.

Novel materials drive progress across applications from energy storage to electronics. Automated characterization of material structures with machine learning methods offers a promising strategy for accelerating this key step in material design. In this work, we introduce an autoregressive language model that performs crystal structure prediction (CSP) from powder diffraction data. The presented model, deCIFer, generates crystal structures in the widely used Crystallographic Information File (CIF) format and can be conditioned on powder X-ray diffraction (PXRD) data. Unlike earlier works that primarily rely on high-level descriptors like composition, deCIFer performs CSP from diffraction data. We train deCIFer on nearly 2.3M unique crystal structures and validate on diverse sets of PXRD patterns for characterizing challenging inorganic crystal systems. Qualitative and quantitative assessments using the residual weighted profile and Wasserstein distance show that deCIFer produces structures that more accurately match the target diffraction data when conditioned, compared to the unconditioned case. Notably, deCIFer can achieve a 94% match rate on unseen data. deCIFer bridges experimental diffraction data with computational CSP, lending itself as a powerful tool for crystal structure characterization and accelerating materials discovery.

In this work, we argue that large language models (LLMs), though trained to predict only the next token, exhibit emergent planning behaviors: $\textbf{their hidden representations encode future outputs beyond the next token}$. Through simple probing, we demonstrate that LLM prompt representations encode global attributes of their entire responses, including $\textit{structural attributes}$ (response length, reasoning steps), $\textit{content attributes}$ (character choices in storywriting, multiple-choice answers at the end of response), and $\textit{behavioral attributes}$ (answer confidence, factual consistency). In addition to identifying response planning, we explore how it scales with model size across tasks and how it evolves during generation. The findings that LLMs plan ahead for the future in their hidden representations suggests potential applications for improving transparency and generation control.

The security issue of large language models (LLMs) has gained significant attention recently, with various defense mechanisms developed to prevent harmful outputs, among which safeguards based on text embedding models serve as a fundamental defense. Through testing, we discover that the distribution of text embedding model outputs is significantly biased with a large mean. Inspired by this observation, we propose novel efficient methods to search for universal magic words that can attack text embedding models. The universal magic words as suffixes can move the embedding of any text towards the bias direction, therefore manipulate the similarity of any text pair and mislead safeguards. By appending magic words to user prompts and requiring LLMs to end answers with magic words, attackers can jailbreak the safeguard. To eradicate this security risk, we also propose defense mechanisms against such attacks, which can correct the biased distribution of text embeddings in a train-free manner.

We propose a new long-context financial benchmark, FailSafeQA, designed to test the robustness and context-awareness of LLMs against six variations in human-interface interactions in LLM-based query-answer systems within finance. We concentrate on two case studies: Query Failure and Context Failure. In the Query Failure scenario, we perturb the original query to vary in domain expertise, completeness, and linguistic accuracy. In the Context Failure case, we simulate the uploads of degraded, irrelevant, and empty documents. We employ the LLM-as-a-Judge methodology with Qwen2.5-72B-Instruct and use fine-grained rating criteria to define and calculate Robustness, Context Grounding, and Compliance scores for 24 off-the-shelf models. The results suggest that although some models excel at mitigating input perturbations, they must balance robust answering with the ability to refrain from hallucinating. Notably, Palmyra-Fin-128k-Instruct, recognized as the most compliant model, maintained strong baseline performance but encountered challenges in sustaining robust predictions in 17% of test cases. On the other hand, the most robust model, OpenAI o3-mini, fabricated information in 41% of tested cases. The results demonstrate that even high-performing models have significant room for improvement and highlight the role of FailSafeQA as a tool for developing LLMs optimized for dependability in financial applications. The dataset is available at: https://huggingface.co/datasets/Writer/FailSafeQA

Reliably generating structured outputs has become a critical capability for modern language model (LM) applications. Constrained decoding has emerged as the dominant technology across sectors for enforcing structured outputs during generation. Despite its growing adoption, little has been done with the systematic evaluation of the behaviors and performance of constrained decoding. Constrained decoding frameworks have standardized around JSON Schema as a structured data format, with most uses guaranteeing constraint compliance given a schema. However, there is poor understanding of the effectiveness of the methods in practice. We present an evaluation framework to assess constrained decoding approaches across three critical dimensions: efficiency in generating constraint-compliant outputs, coverage of diverse constraint types, and quality of the generated outputs. To facilitate this evaluation, we introduce JSONSchemaBench, a benchmark for constrained decoding comprising 10K real-world JSON schemas that encompass a wide range of constraints with varying complexity. We pair the benchmark with the existing official JSON Schema Test Suite and evaluate six state-of-the-art constrained decoding frameworks, including Guidance, Outlines, Llamacpp, XGrammar, OpenAI, and Gemini. Through extensive experiments, we gain insights into the capabilities and limitations of constrained decoding on structured generation with real-world JSON schemas. Our work provides actionable insights for improving constrained decoding frameworks and structured generation tasks, setting a new standard for evaluating constrained decoding and structured generation. We release JSONSchemaBench at https://github.com/guidance-ai/jsonschemabench

Decentralized finance (DeFi) applications depend on accurate price oracles to ensure secure transactions, yet these oracles are highly vulnerable to manipulation, enabling attackers to exploit smart contract vulnerabilities for unfair asset valuation and financial gain. Detecting such manipulations traditionally relies on the manual effort of experienced experts, presenting significant challenges. In this paper, we propose a novel LLM-driven framework that automates the detection of price oracle manipulations by leveraging the complementary strengths of different LLM models (LLMs). Our approach begins with domain-specific knowledge extraction, where an LLM model synthesizes precise insights about price oracle vulnerabilities from top-tier academic papers, eliminating the need for profound expertise from developers or auditors. This knowledge forms the foundation for a second LLM model to generate structured, context-aware chain of thought prompts, which guide a third LLM model in accurately identifying manipulation patterns in smart contracts. We validate the effectiveness of framework through experiments on 60 known vulnerabilities from 46 real-world DeFi attacks or projects spanning 2021 to 2023. The best performing combination of LLMs (Haiku-Haiku-4o-mini) identified by AiRacleX demonstrate a 2.58-times improvement in recall (0.667 vs 0.259) compared to the state-of-the-art tool GPTScan, while maintaining comparable precision. Furthermore, our framework demonstrates the feasibility of replacing commercial models with open-source alternatives, enhancing privacy and security for developers.