Evaluating how Large Language Models (LLMs) handle complex, specialized knowledge remains a critical challenge. We address this through the lens of climate change by introducing CLINB, a benchmark that assesses models on open-ended, grounded, multimodal question answering tasks with clear requirements for knowledge quality and evidential support. CLINB relies on a dataset of real users' questions and evaluation rubrics curated by leading climate scientists. We implement and validate a model-based evaluation process and evaluate several frontier models. Our findings reveal a critical dichotomy. Frontier models demonstrate remarkable knowledge synthesis capabilities, often exhibiting PhD-level understanding and presentation quality. They outperform "hybrid" answers curated by domain experts assisted by weaker models. However, this performance is countered by failures in grounding. The quality of evidence varies, with substantial hallucination rates for references and images. We argue that bridging this gap between knowledge synthesis and verifiable attribution is essential for the deployment of AI in scientific workflows and that reliable, interpretable benchmarks like CLINB are needed to progress towards building trustworthy AI systems.

Traditional fuzzy matching often fails when searching for quotes that are semantically identical but syntactically different across documents-a common issue when aligning official written records with speech-to-text transcripts. We introduce TimeStampEval, a benchmark for retrieving precise millisecond timestamps from long transcripts given non-verbatim quotes. Our simple two-stage method dramatically improves retrieval accuracy while cutting inference costs by over 90%. The motivating use case is an automated long-form podcast that assembles Congressional Record clips into AI-hosted narration. The technical challenge: given a sentence-timestamped transcript and a target quote that may differ due to transcription or editorial drift, return exact start and end boundaries. Standard algorithms handle verbatim text but break under fuzzier variants. Evaluating six modern LLMs on a 2,800-sentence (120k-token) transcript revealed four key findings. (1) Prompt design matters more than model choice: placing the query before the transcript and using compact formatting improved accuracy by 3-20 points while reducing token count by 30-40%. (2) Off-by-one errors form a distinct category, showing models understand the task but misplace boundaries. (3) A modest reasoning budget (600-850 tokens) raises accuracy from 37% to 77% for weak setups and to above 90% for strong ones. (4) Our "Assisted Fuzzy" approach-RapidFuzz pre-filtering followed by LLM verification on short snippets-improves fuzzy match accuracy by up to 50 points while halving latency and reducing cost per correct result by up to 96%. Extended tests on ten transcripts (50k-900k tokens, 1989-2025) confirm robustness to transcript length, vocabulary drift, and domain change, maintaining 95-100% rejection accuracy for absent targets.

This research examines the potential of datasets generated by Large Language Models (LLMs) to support Natural Language Processing (NLP) tasks, aiming to overcome challenges related to data acquisition and privacy concerns associated with real-world data. Focusing on negative valence text, a critical component of sentiment analysis, we explore the use of LLM-generated synthetic news headlines as an alternative to real-world data. A specialized corpus of negative news headlines was created using tailored prompts to capture diverse negative sentiments across various societal domains. The synthetic headlines were validated by expert review and further analyzed in embedding space to assess their alignment with real-world negative news in terms of content, tone, length, and style. Key metrics such as correlation with real headlines, perplexity, coherence, and realism were evaluated. The synthetic dataset was benchmarked against two sets of real news headlines using evaluations including the Comparative Perplexity Test, Comparative Readability Test, Comparative POS Profiling, BERTScore, and Comparative Semantic Similarity. Results show the generated headlines match real headlines with the only marked divergence being in the proper noun score of the POS profile test.

OpenAI (2025a) showed that training against a chain of thought (CoT) monitor can cause obfuscated CoTs, which contain bad behavior the monitor cannot detect. They proposed to keep CoTs monitorable by training only against output monitors that do not have access to CoT. We show that such training can still cause obfuscated CoTs via two mechanisms. First, when a model is trained to produce a safe-looking output, that model may generalize to making its CoTs look safe. Second, since later tokens are conditioned on earlier ones, safe-looking CoTs may increase the likelihood of safe outputs, causing safe-looking CoTs to be reinforced. We introduce two mitigations to address these two issues, which achieve a Pareto improvement in terms of monitorability and task performance compared to regular training. Our work implies that preserving CoT monitorabil-ity is more difficult than previously thought; we suggest practical guidelines for AI developers to maintain monitorable CoTs. Reasoning models, which generate a CoT before their final answer, dominate mathematics and coding benchmarks (Guo et al., 2025; OpenAI, 2025b; Qwen Team, 2025). Korbak et al. (2025) argue that AI developers should aim to maintain the ability to detect model misbehavior by analyzing CoTs ("CoT monitorability"), claiming that it offers a "unique opportunity for AI safety". Baker et al. (2025) show that training frontier models against a CoT monitor can cause their CoTs to become obfuscated: unwanted behaviors persist in a way that is undetectable by the monitor. To address this, they recommend against directly optimizing the CoT. However, we show that even training against a monitor that sees only the final output (and not the CoT) can still cause obfuscated CoT. We refer to this effect as feedback spillover, and show that it can occur for two reasons:Figure 1: Feedback spillover: training against an output monitor obfuscates the CoT. 1 To mitigate parametric feedback spillover, we generate the CoT and output using two distinct models, which we refer to as the "Mind" and "Face" respectively, as proposed by Kokotajlo & Demski (2024).

Large language models (LLMs) show promise for personalized financial recommendations but are hampered by context limits, hallucinations, and a lack of behavioral grounding. Our prior work, FLARKO, embedded structured knowledge graphs (KGs) in LLM prompts to align advice with user behavior and market data. This paper introduces RAG-FLARKO, a retrieval-augmented extension to FLARKO, that overcomes scalability and relevance challenges using multi-stage and parallel KG retrieval processes. Our method first retrieves behaviorally relevant entities from a user's transaction KG and then uses this context to filter temporally consistent signals from a market KG, constructing a compact, grounded subgraph for the LLM. This pipeline reduces context overhead and sharpens the model's focus on relevant information. Empirical evaluation on a real-world financial transaction dataset demonstrates that RAG-FLARKO significantly enhances recommendation quality. Notably, our framework enables smaller, more efficient models to achieve high performance in both profitability and behavioral alignment, presenting a viable path for deploying grounded financial AI in resource-constrained environments.

A long-standing goal in both industry and academia is to develop an LLM inference platform that is portable across hardware architectures, eliminates the need for low-level hand-tuning, and still delivers best-in-class efficiency. In this work, we demonstrate that portable, efficient cross-platform LLM inference is indeed possible and share our experience. We develop a state-of-the-art paged attention kernel, the core performance-critical component of many LLM deployments, that builds exclusively on the domain-specific just-in-time compiled language Triton to achieve state-of-the-art performance on both NVIDIA and AMD GPUs. We describe our high-level approach, the key algorithmic and system-level improvements, the parameter auto-tuning required to unlock efficiency, and the integrations into a popular inference server that are necessary to bring the performance of a generic Triton attention kernel from 19.7% of the state-of-the-art to 105.9%. Our results highlight how open-source domain-specific languages can be leveraged to unlock model portability across different GPU vendors.

An important aspect subtending language understanding and production is the ability to independently encode positional and symbolic information of the words within a sentence. In Transformers, positional information is typically encoded using Positional Encodings (PEs). One such popular PE, namely Rotary PE (RoPE), has been widely used due to its empirical success. Recently, it has been argued that part of RoPE's success emerges from its ability to encode robust positional and semantic information using large and small frequencies, respectively. In this work, we perform a deeper dive into the positional versus symbolic dichotomy of attention heads behavior, both at the theoretical and empirical level. We provide general definitions of what it means for a head to behave positionally or symbolically, prove that these are two mutually exclusive behaviors and develop a metric to quantify them. We apply our framework to analyze Transformer-based LLMs using RoPE and find that all heads exhibit a strong correspondence between behavior and frequency use. Finally, we introduce canonical tasks designed to be either purely positional or symbolic, and demonstrate that the Transformer performance causally relates to the ability of attention heads to leverage the appropriate frequencies. In particular, we show that we can control the Transformer performance by controlling which frequencies the attention heads can access. Altogether, our work provides a detailed understanding of RoPE, and how its properties relate to model behavior.

Recent advances in large language models (LLMs) have accelerated research on automated optimization modeling. While real-world decision-making is inherently uncertain, most existing work has focused on deterministic optimization with known parameters, leaving the application of LLMs in uncertain settings largely unexplored. To that end, we propose the DAOpt framework including a new dataset OptU, a multi-agent decision-making module, and a simulation environment for evaluating LLMs with a focus on out-of-sample feasibility and robustness. Additionally, we enhance LLMs' modeling capabilities by incorporating few-shot learning with domain knowledge from stochastic and robust optimization.

The current standard architecture of Large Language Models (LLMs) with QKV self-attention is briefly summarized, including the architecture of a typical Transformer. Scaling laws for compute (flops) and memory (parameters plus data) are given, along with their present (2025) rough cost estimates for the parameters of present LLMs of various scales, including discussion of whether DeepSeek should be viewed as a special case. Nothing here is new, but this material seems not otherwise readily available in summary form.

Self-consistency (SC) is a widely used test-time inference technique for improving performance in chain-of-thought reasoning. It involves generating multiple responses, or samples from a large language model (LLM) and selecting the most frequent answer. This procedure can naturally be viewed as a majority vote or empirical mode estimation. Despite its effectiveness, SC is prohibitively expensive at scale when naively applied to datasets, and it lacks a unified theoretical treatment of sample efficiency and scaling behavior. In this paper, we provide the first comprehensive analysis of SC's scaling behavior and its variants, drawing on mode estimation and voting theory. We derive and empirically validate power law scaling for self-consistency across datasets, and analyze the sample efficiency for fixed-allocation and dynamic-allocation sampling schemes. From these insights, we introduce Blend-ASC, a novel variant of self-consistency that dynamically allocates samples to questions during inference, achieving state-of-the-art sample efficiency. Our approach uses 6.8x fewer samples than vanilla SC on average, outperforming both fixed- and dynamic-allocation SC baselines, thereby demonstrating the superiority of our approach in terms of efficiency. In contrast to existing variants, Blend-ASC is hyperparameter-free and can fit an arbitrary sample budget, ensuring it can be easily applied to any self-consistency application.