Large Language Models (LLMs) have recently emerged as planners for language-instructed agents, generating sequences of actions to accomplish natural language tasks. However, their reliability remains a challenge, especially in long-horizon tasks, since they often produce overconfident yet wrong outputs. Conformal Prediction (CP) has been leveraged to address this issue by wrapping LLM outputs into prediction sets that contain the correct action with a user-defined confidence. When the prediction set is a singleton, the planner executes that action; otherwise, it requests help from a user. This has led to LLM-based planners that can ensure plan correctness with a user-defined probability. However, as LLMs are trained in an uncertainty-agnostic manner, without awareness of prediction sets, they tend to produce unnecessarily large sets, particularly at higher confidence levels, resulting in frequent human interventions limiting autonomous deployment. To address this, we introduce CoFineLLM (Conformal Finetuning for LLMs), the first CP-aware fine-tuning framework for LLM-based planners that explicitly reduces prediction-set size and, in turn, the need for user interventions. We evaluate our approach on multiple language-instructed robot planning problems and show consistent improvements over uncertainty-aware and uncertainty-agnostic finetuning baselines in terms of prediction-set size, and help rates. Finally, we demonstrate robustness of our method to out-of-distribution scenarios in hardware experiments.

Large language models (LLMs) are increasingly used as behavioral proxies for self-interested travelers in agent-based traffic models. Although more flexible and generalizable than conventional models, the practical use of these approaches remains limited by scalability due to the cost of calling one LLM for every traveler. Moreover, it has been found that LLM agents often make opaque choices and produce unstable day-to-day dynamics. To address these challenges, we propose to model each homogeneous traveler group facing the same decision context with a single representative LLM agent who behaves like the population's average, maintaining and updating a mixed strategy over routes that coincides with the group's aggregate flow proportions. Each day, the LLM reviews the travel experience and flags routes with positive reinforcement that they hope to use more often, and an interpretable update rule then converts this judgment into strategy adjustments using a tunable (progressively decaying) step size. The representative-agent design improves scalability, while the separation of reasoning from updating clarifies the decision logic while stabilizing learning. In classic traffic assignment settings, we find that the proposed approach converges rapidly to the user equilibrium. In richer settings with income heterogeneity, multi-criteria costs, and multi-modal choices, the generated dynamics remain stable and interpretable, reproducing plausible behavioral patterns well-documented in psychology and economics, for example, the decoy effect in toll versus non-toll road selection, and higher willingness-to-pay for convenience among higher-income travelers when choosing between driving, transit, and park-and-ride options.

The rapid progress of large language models (LLMs) has advanced numerous applications, yet efficient single-batch inference remains vital for on-device intelligence. While FPGAs offer fine-grained data control and high energy efficiency, recent GPU optimizations have narrowed their advantage, especially under arithmetic-based computation. To overcome this, we leverage FPGAs' abundant on-chip memory to shift LLM inference from arithmetic- to memory-based computation through table lookups. We present LUT-LLM, the first FPGA accelerator enabling 1B+ LLM inference via vector-quantized memory operations. Our analysis identifies activation-weight co-quantization as the most effective scheme, supported by (1) bandwidth-aware parallel centroid search, (2) efficient 2D table lookups, and (3) a spatial-temporal hybrid design minimizing data caching. Implemented on an AMD V80 FPGA for a customized Qwen 3 1.7B model, LUT-LLM achieves 1.66x lower latency than AMD MI210 and 1.72x higher energy efficiency than NVIDIA A100, scaling to 32B models with 2.16x efficiency gain over A100.

Large language models (LLMs) can internally distinguish between evaluation and deployment contexts, a behaviour known as \emph{evaluation awareness}. This undermines AI safety evaluations, as models may conceal dangerous capabilities during testing. Prior work demonstrated this in a single $70$B model, but the scaling relationship across model sizes remains unknown. We investigate evaluation awareness across $15$ models scaling from $0.27$B to $70$B parameters from four families using linear probing on steering vector activations. Our results reveal a clear power-law scaling: evaluation awareness increases predictably with model size. This scaling law enables forecasting deceptive behavior in future larger models and guides the design of scale-aware evaluation strategies for AI safety. A link to the implementation of this paper can be found at https://anonymous.4open.science/r/evaluation-awareness-scaling-laws/README.md.

Shaping inclusive representations that embrace diversity and ensure fair participation and reflections of values is at the core of many conversation-based models. However, many existing methods rely on surface inclusion using mention of user demographics or behavioral attributes of social groups. Such methods overlook the nuanced, implicit expression of opinion embedded in conversations. Furthermore, the over-reliance on overt cues can exacerbate misalignment and reinforce harmful or stereotypical representations in model outputs. Thus, we took a step back and recognized that equitable inclusion needs to account for the implicit expression of opinion and use the stance of responses to validate the normative alignment. This study aims to evaluate how opinions are represented in NLP or computational models by introducing an alignment evaluation framework that foregrounds implicit, often overlooked conversations and evaluates the normative social views and discourse. Our approach models the stance of responses as a proxy for the underlying opinion, enabling a considerate and reflective representation of diverse social viewpoints. We evaluate the framework using both (i) positive-unlabeled (PU) online learning with base classifiers, and (ii) instruction-tuned language models to assess post-training alignment. Through this, we provide a principled and structured lens on how implicit opinions are (mis)represented and offer a pathway toward more inclusive model behavior.

When an LLM learns a new fact during finetuning (e.g., new movie releases, newly elected pope, etc.), where does this information go? Are entities enriched with relation information, or do models recall information just-in-time before a prediction? Or, are ``all of the above'' true with LLMs implementing multiple redundant heuristics? Existing localization approaches (e.g., activation patching) are ill-suited for this analysis because they usually \textit{replace} parts of the residual stream, thus overriding previous information. To fill this gap, we propose \emph{dynamic weight grafting}, a technique that selectively grafts weights from a finetuned model onto a pretrained model. Using this technique, we show two separate pathways for retrieving finetuned relation information: 1) ``enriching" the residual stream with relation information while processing the tokens that correspond to an entity (e.g., ``Zendaya'' in ``Zendaya co-starred with John David Washington'') and 2) ``recalling" this information at the final token position before generating a target fact. In some cases, models need information from both of these pathways to correctly generate finetuned facts while, in other cases, either the ``enrichment" or ``recall" pathway alone is sufficient. We localize the ``recall'' pathway to model components -- finding that ``recall" occurs via both task-specific attention mechanisms and an entity-specific extraction step in the feedforward networks of the final layers before the target prediction. By targeting model components and parameters, as opposed to just activations, we are able to understand the \textit{mechanisms} by which finetuned knowledge is retrieved during generation.

This paper presents a Large Language Model (LLM) based conversational agent system designed to enhance human-machine collaboration in Machine Learning Operations (MLOps). We introduce the Swarm Agent, an extensible architecture that integrates specialized agents to create and manage ML workflows through natural language interactions. The system leverages a hierarchical, modular design incorporating a KubeFlow Pipelines (KFP) Agent for ML pipeline orchestration, a MinIO Agent for data management, and a Retrieval-Augmented Generation (RAG) Agent for domain-specific knowledge integration. Through iterative reasoning loops and context-aware processing, the system enables users with varying technical backgrounds to discover, execute, and monitor ML pipelines; manage datasets and artifacts; and access relevant documentation, all via intuitive conversational interfaces. Our approach addresses the accessibility gap in complex MLOps platforms like Kubeflow, making advanced ML tools broadly accessible while maintaining the flexibility to extend to other platforms. The paper describes the architecture, implementation details, and demonstrates how this conversational MLOps assistant reduces complexity and lowers barriers to entry for users across diverse technical skill levels.

This paper presents a simple, effective, and cost-efficient strategy to improve LLM performance by scaling test-time compute. Our strategy builds upon the repeated-sampling-then-voting framework, with a novel twist: incorporating multiple models, even weaker ones, to leverage their complementary strengths that potentially arise from diverse training data and paradigms. By using consistency as a signal, our strategy dynamically switches between models. Theoretical analysis highlights the efficiency and performance advantages of our strategy. Extensive experiments on six datasets demonstrate that our strategy not only outperforms self-consistency and state-of-the-art multi-agent debate approaches, but also significantly reduces inference costs. Additionally, ModelSwitch requires only a few comparable LLMs to achieve optimal performance and can be extended with verification methods, demonstrating the potential of leveraging multiple LLMs in the generation-verification paradigm.

Numerous multimodal misinformation benchmarks exhibit bias toward specific modalities, allowing detectors to make predictions based solely on one modality. While previous research has quantified bias at the dataset level or manually identified spurious correlations between modalities and labels, these approaches lack meaningful insights at the sample level and struggle to scale to the vast amount of online information. In this paper, we investigate the design for automated recognition of modality bias at the sample level. Specifically, we propose three bias quantification methods based on theories/views of different levels of granularity: 1) a coarse-grained evaluation of modality benefit; 2) a medium-grained quantification of information flow; and 3) a fine-grained causality analysis. T o verify the effectiveness, we conduct a human evaluation on two popular benchmarks. Experimental results reveal three interesting findings that provide potential direction toward future research: 1) Ensembling multiple views is crucial for reliable automated analysis; 2) Automated analysis is prone to detector-induced fluctuations; and 3) Different views produce a higher agreement on modality-balanced samples but diverge on biased ones.

Alignment has greatly improved large language models (LLMs)' output quality at the cost of diversity, yielding highly similar outputs across generations. We propose Base-Aligned Model Collaboration (BACo), an inference-time token-level model collaboration framework that dynamically combines a base LLM with its aligned counterpart to optimize diversity and quality. Inspired by prior work (Fei et al., 2025), BACo employs routing strategies that determine, at each token, from which model to decode based on next-token prediction uncertainty and predicted contents' semantic role. Prior diversity-promoting methods, such as retraining, prompt engineering, and multi-sampling methods, improve diversity but often degrade quality or require costly decoding or post-training. In contrast, BACo achieves both high diversity and quality post hoc within a single pass, while offering strong controllability. We explore a family of routing strategies, across three open-ended generation tasks and 13 metrics covering diversity and quality, BACo consistently surpasses state-of-the-art inference-time baselines. With our best router, BACo achieves a 21.3% joint improvement in diversity and quality. Human evaluations also mirror these improvements. The results suggest that collaboration between base and aligned models can optimize and control diversity and quality.