Autonomous agents powered by large language models (LLMs) enable novel use cases in domains where responsible action is increasingly important. Yet the inherent unpredictability of LLMs raises safety concerns about agent reliability. In this work, we explore agent behaviour in a toy, game-theoretic environment based on a variation of the Iterated Prisoner's Dilemma. We introduce a strategy-modification method-independent of both the game and the prompt-by steering the residual stream with interpretable features extracted from a sparse autoencoder latent space. Steering with the good-faith negotiation feature lowers the average defection probability by 28 percentage points. We also identify feasible steering ranges for several open-source LLM agents. Finally, we hypothesise that game-theoretic evaluation of LLM agents, combined with representation-steering alignment, can generalise to real-world applications on end-user devices and embodied platforms.
Large language models (LLMs) have opened new opportunities for automated mobile app exploration, an important and challenging problem that used to suffer from the difficulty of generating meaningful UI interactions. However, existing LLM-based exploration approaches rely heavily on LLMs to generate actions in almost every step, leading to a huge cost of token fees and computational resources. We argue that such extensive usage of LLMs is neither necessary nor effective, since many actions during exploration do not require, or may even be biased by the abilities of LLMs. Further, based on the insight that a precise and compact knowledge plays the central role for effective exploration, we introduce LLM-Explorer, a new exploration agent designed for efficiency and affordability. LLM-Explorer uses LLMs primarily for maintaining the knowledge instead of generating actions, and knowledge is used to guide action generation in a LLM-less manner. Based on a comparison with 5 strong baselines on 20 typical apps, LLM-Explorer was able to achieve the fastest and highest coverage among all automated app explorers, with over 148x lower cost than the state-of-the-art LLM-based approach.
We present SimpleStories, a large synthetic story dataset in simple language, consisting of 2 million samples each in English and Japanese. Through parameterizing prompts at multiple levels of abstraction, we achieve control over story characteristics at scale, inducing syntactic and semantic diversity. Ablations on a newly trained model suite show improved sample efficiency and model interpretability compared to the TinyStories dataset. We open-source all constituent parts of model creation, hoping to enable novel ways to study the end-to-end training process. As a byproduct, we move the frontier regarding the fewest-parameter language model that outputs grammatical natural language.
Large language models (LLMs) are increasingly used as automatic evaluators in applications such as benchmarking, reward modeling, and self-refinement. Prior work highlights a potential self-preference bias where LLMs favor their own generated responses, a tendency often intensifying with model size and capability. This raises a critical question: Is self-preference detrimental, or does it simply reflect objectively superior outputs from more capable models? Disentangling these has been challenging due to the usage of subjective tasks in previous studies. To address this, we investigate self-preference using verifiable benchmarks (mathematical reasoning, factual knowledge, code generation) that allow objective ground-truth assessment. This enables us to distinguish harmful self-preference (favoring objectively worse responses) from legitimate self-preference (favoring genuinely superior ones). We conduct large-scale experiments under controlled evaluation conditions across diverse model families (e.g., Llama, Qwen, Gemma, Mistral, Phi, GPT, DeepSeek). Our findings reveal three key insights: (1) Better generators are better judges -- LLM evaluators' accuracy strongly correlates with their task performance, and much of the self-preference in capable models is legitimate. (2) Harmful self-preference persists, particularly when evaluator models perform poorly as generators on specific task instances. Stronger models exhibit more pronounced harmful bias when they err, though such incorrect generations are less frequent. (3) Inference-time scaling strategies, such as generating a long Chain-of-Thought before evaluation, effectively reduce the harmful self-preference. These results provide a more nuanced understanding of LLM-based evaluation and practical insights for improving its reliability.
Test-Time Scaling (TTS) refers to approaches that improve reasoning performance by allocating extra computation during inference, without altering the model's parameters. While existing TTS methods operate in a discrete token space by generating more intermediate steps, recent studies in Coconut and SoftCoT have demonstrated that thinking in the continuous latent space can further enhance the reasoning performance. Such latent thoughts encode informative thinking without the information loss associated with autoregressive token generation, sparking increased interest in continuous-space reasoning. Unlike discrete decoding, where repeated sampling enables exploring diverse reasoning paths, latent representations in continuous space are fixed for a given input, which limits diverse exploration, as all decoded paths originate from the same latent thought. To overcome this limitation, we introduce SoftCoT++ to extend SoftCoT to the Test-Time Scaling paradigm by enabling diverse exploration of thinking paths. Specifically, we perturb latent thoughts via multiple specialized initial tokens and apply contrastive learning to promote diversity among soft thought representations. Experiments across five reasoning benchmarks and two distinct LLM architectures demonstrate that SoftCoT++ significantly boosts SoftCoT and also outperforms SoftCoT with self-consistency scaling. Moreover, it shows strong compatibility with conventional scaling techniques such as self-consistency. Source code is available at https://github.com/xuyige/SoftCoT.
Video Comment Art enhances user engagement by providing creative content that conveys humor, satire, or emotional resonance, requiring a nuanced and comprehensive grasp of cultural and contextual subtleties. Although Multimodal Large Language Models (MLLMs) and Chain-of-Thought (CoT) have demonstrated strong reasoning abilities in STEM tasks (e.g. mathematics and coding), they still struggle to generate creative expressions such as resonant jokes and insightful satire. Moreover, existing benchmarks are constrained by their limited modalities and insufficient categories, hindering the exploration of comprehensive creativity in video-based Comment Art creation. To address these limitations, we introduce GODBench, a novel benchmark that integrates video and text modalities to systematically evaluate MLLMs' abilities to compose Comment Art. Furthermore, inspired by the propagation patterns of waves in physics, we propose Ripple of Thought (RoT), a multi-step reasoning framework designed to enhance the creativity of MLLMs. Extensive experiments reveal that existing MLLMs and CoT methods still face significant challenges in understanding and generating creative video comments. In contrast, RoT provides an effective approach to improve creative composing, highlighting its potential to drive meaningful advancements in MLLM-based creativity. GODBench is publicly available at https://github.com/stan-lei/GODBench-ACL2025.
AI spans from large language models to tiny models running on microcontrollers (MCUs). Extremely memory-efficient model architectures are decisive to fit within an MCU's tiny memory budget e.g., 128kB of RAM. However, inference latency must remain small to fit real-time constraints. An approach to tackle this is patch-based fusion, which aims to optimize data flows across neural network layers. In this paper, we introduce msf-CNN, a novel technique that efficiently finds optimal fusion settings for convolutional neural networks (CNNs) by walking through the fusion solution space represented as a directed acyclic graph. Compared to previous work on CNN fusion for MCUs, msf-CNN identifies a wider set of solutions. We published an implementation of msf-CNN running on various microcontrollers (ARM Cortex-M, RISC-V, ESP32). We show that msf-CNN can achieve inference using 50% less RAM compared to the prior art (MCUNetV2 and StreamNet). We thus demonstrate how msf-CNN offers additional flexibility for system designers.
This paper introduces a novel dataset and evaluation benchmark designed to assess and improve small language models deployable on edge devices, with a focus on user profiling from multi-session natural language interactions in smart home environments. At the core of the dataset are structured user profiles, each defined by a set of routines - context-triggered, repeatable patterns of behavior that govern how users interact with their home systems. Using these profiles as input, a large language model (LLM) generates corresponding interaction sessions that simulate realistic, diverse, and context-aware dialogues between users and their devices. The primary task supported by this dataset is profile reconstruction: inferring user routines and preferences solely from interactions history. To assess how well current models can perform this task under realistic conditions, we benchmarked several state-of-the-art compact language models and compared their performance against large foundation models. Our results show that while small models demonstrate some capability in reconstructing profiles, they still fall significantly short of large models in accurately capturing user behavior. This performance gap poses a major challenge - particularly because on-device processing offers critical advantages, such as preserving user privacy, minimizing latency, and enabling personalized experiences without reliance on the cloud. By providing a realistic, structured testbed for developing and evaluating behavioral modeling under these constraints, our dataset represents a key step toward enabling intelligent, privacy-respecting AI systems that learn and adapt directly on user-owned devices.
Counterfactual learning is emerging as an important paradigm, rooted in causality, which promises to alleviate common issues of graph neural networks (GNNs), such as fairness and interpretability. However, as in many real-world application domains where conducting randomized controlled trials is impractical, one has to rely on available observational (factual) data to detect counterfactuals. In this paper, we introduce and tackle the problem of searching for counterfactual evidences for the GNN-based node classification task. A counterfactual evidence is a pair of nodes such that, regardless they exhibit great similarity both in the features and in their neighborhood subgraph structures, they are classified differently by the GNN. We develop effective and efficient search algorithms and a novel indexing solution that leverages both node features and structural information to identify counterfactual evidences, and generalizes beyond any specific GNN. Through various downstream applications, we demonstrate the potential of counterfactual evidences to enhance fairness and accuracy of GNNs.
When the distribution of the data used to train a classifier differs from that of the test data, i.e., under dataset shift, well-established routines for calibrating the decision scores of the classifier, estimating the proportion of positives in a test sample, or estimating the accuracy of the classifier, become particularly challenging. This paper investigates the interconnections among three fundamental problems, calibration, quantification, and classifier accuracy prediction, under dataset shift conditions. Specifically, we prove their equivalence through mutual reduction, i.e., we show that access to an oracle for any one of these tasks enables the resolution of the other two. Based on these proofs, we propose new methods for each problem based on direct adaptations of well-established methods borrowed from the other disciplines. Our results show such methods are often competitive, and sometimes even surpass the performance of dedicated approaches from each discipline. The main goal of this paper is to fostering cross-fertilization among these research areas, encouraging the development of unified approaches and promoting synergies across the fields.
Current invasive assistive technologies are designed to infer high-dimensional motor control signals from severely paralyzed patients. However, they face significant challenges, including public acceptance, limited longevity, and barriers to commercialization. Meanwhile, noninvasive alternatives often rely on artifact-prone signals, require lengthy user training, and struggle to deliver robust high-dimensional control for dexterous tasks. To address these issues, this study introduces a novel human-centered multimodal AI approach as intelligent compensatory mechanisms for lost motor functions that could potentially enable patients with severe paralysis to control high-dimensional assistive devices, such as dexterous robotic arms, using limited and noninvasive inputs. In contrast to the current state-of-the-art (SoTA) noninvasive approaches, our context-aware, multimodal shared-autonomy framework integrates deep reinforcement learning algorithms to blend limited low-dimensional user input with real-time environmental perception, enabling adaptive, dynamic, and intelligent interpretation of human intent for complex dexterous manipulation tasks, such as pick-and-place. The results from our ARAS (Adaptive Reinforcement learning for Amplification of limited inputs in Shared autonomy) trained with synthetic users over 50,000 computer simulation episodes demonstrated the first successful implementation of the proposed closed-loop human-in-the-loop paradigm, outperforming the SoTA shared autonomy algorithms. Following a zero-shot sim-to-real transfer, ARAS was evaluated on 23 human subjects, demonstrating high accuracy in dynamic intent detection and smooth, stable 3D trajectory control for dexterous pick-and-place tasks. ARAS user study achieved a high task success rate of 92.88%, with short completion times comparable to those of SoTA invasive assistive technologies.
Style Transfer with Inference-Time Optimisation (ST-ITO) is a recent approach for transferring the applied effects of a reference audio to a raw audio track. It optimises the effect parameters to minimise the distance between the style embeddings of the processed audio and the reference. However, this method treats all possible configurations equally and relies solely on the embedding space, which can lead to unrealistic or biased results. We address this pitfall by introducing a Gaussian prior derived from a vocal preset dataset, DiffVox, over the parameter space. The resulting optimisation is equivalent to maximum-a-posteriori estimation. Evaluations on vocal effects transfer on the MedleyDB dataset show significant improvements across metrics compared to baselines, including a blind audio effects estimator, nearest-neighbour approaches, and uncalibrated ST-ITO. The proposed calibration reduces parameter mean squared error by up to 33% and matches the reference style better. Subjective evaluations with 16 participants confirm our method's superiority, especially in limited data regimes. This work demonstrates how incorporating prior knowledge in inference time enhances audio effects transfer, paving the way for more effective and realistic audio processing systems.
Meta-World is widely used for evaluating multi-task and meta-reinforcement learning agents, which are challenged to master diverse skills simultaneously. Since its introduction however, there have been numerous undocumented changes which inhibit a fair comparison of algorithms. This work strives to disambiguate these results from the literature, while also leveraging the past versions of Meta-World to provide insights into multi-task and meta-reinforcement learning benchmark design. Through this process we release a new open-source version of Meta-World (https://github.com/Farama-Foundation/Metaworld/) that has full reproducibility of past results, is more technically ergonomic, and gives users more control over the tasks that are included in a task set.
Large language models have demonstrated impressive reasoning capabilities but are inherently limited by their knowledge reservoir. Retrieval-augmented reasoning mitigates this limitation by allowing LLMs to query external resources, but existing methods often retrieve irrelevant or noisy information, hindering accurate reasoning. In this paper, we propose AutoRefine, a reinforcement learning post-training framework that adopts a new ``search-and-refine-during-think'' paradigm. AutoRefine introduces explicit knowledge refinement steps between successive search calls, enabling the model to iteratively filter, distill, and organize evidence before generating an answer. Furthermore, we incorporate tailored retrieval-specific rewards alongside answer correctness rewards using group relative policy optimization. Experiments on single-hop and multi-hop QA benchmarks demonstrate that AutoRefine significantly outperforms existing approaches, particularly in complex, multi-hop reasoning scenarios. Detailed analysis shows that AutoRefine issues frequent, higher-quality searches and synthesizes evidence effectively.
Metaphorical imagination, the ability to connect seemingly unrelated concepts, is fundamental to human cognition and communication. While understanding linguistic metaphors has advanced significantly, grasping multimodal metaphors, such as those found in internet memes, presents unique challenges due to their unconventional expressions and implied meanings. Existing methods for multimodal metaphor identification often struggle to bridge the gap between literal and figurative interpretations. Additionally, generative approaches that utilize large language models or text-to-image models, while promising, suffer from high computational costs. This paper introduces \textbf{C}oncept \textbf{D}rift \textbf{G}uided \textbf{L}ayerNorm \textbf{T}uning (\textbf{CDGLT}), a novel and training-efficient framework for multimodal metaphor identification. CDGLT incorporates two key innovations: (1) Concept Drift, a mechanism that leverages Spherical Linear Interpolation (SLERP) of cross-modal embeddings from a CLIP encoder to generate a new, divergent concept embedding. This drifted concept helps to alleviate the gap between literal features and the figurative task. (2) A prompt construction strategy, that adapts the method of feature extraction and fusion using pre-trained language models for the multimodal metaphor identification task. CDGLT achieves state-of-the-art performance on the MET-Meme benchmark while significantly reducing training costs compared to existing generative methods. Ablation studies demonstrate the effectiveness of both Concept Drift and our adapted LN Tuning approach. Our method represents a significant step towards efficient and accurate multimodal metaphor understanding. The code is available: \href{https://github.com/Qianvenh/CDGLT}{https://github.com/Qianvenh/CDGLT}.
While scaling the length of responses at test-time has been shown to markedly improve the reasoning abilities and performance of large language models (LLMs), it often results in verbose outputs and increases inference cost. Prior approaches for efficient test-time scaling, typically using universal budget constraints or query-level length optimization, do not leverage historical information from previous encounters with the same problem during training. We hypothesize that this limits their ability to progressively make solutions more concise over time. To address this, we present History-Aware Policy Optimization (HAPO), which keeps track of a history state (e.g., the minimum length over previously generated correct responses) for each problem. HAPO employs a novel length reward function based on this history state to incentivize the discovery of correct solutions that are more concise than those previously found. Crucially, this reward structure avoids overly penalizing shorter incorrect responses with the goal of facilitating exploration towards more efficient solutions. By combining this length reward with a correctness reward, HAPO jointly optimizes for correctness and efficiency. We use HAPO to train DeepSeek-R1-Distill-Qwen-1.5B, DeepScaleR-1.5B-Preview, and Qwen-2.5-1.5B-Instruct, and evaluate HAPO on several math benchmarks that span various difficulty levels. Experiment results demonstrate that HAPO effectively induces LLMs' concise reasoning abilities, producing length reductions of 33-59% with accuracy drops of only 2-5%.
Generative AI systems can help spread information but also misinformation and biases, potentially undermining the UN Sustainable Development Goals (SDGs). Explainable AI (XAI) aims to reveal the inner workings of AI systems and expose misbehaviours or biases. However, current XAI tools, built for simpler models, struggle to handle the non-numerical nature of large language models (LLMs). This paper examines the effectiveness of global XAI methods, such as rule-extraction algorithms and SHAP, in detecting bias in LLMs. To do so, we first show a text-to-ordinal mapping strategy to convert non-numerical inputs/outputs into numerical features, enabling these tools to identify (some) misinformation-related biases in LLM-generated content. Then, we inject non-linear biases of varying complexity (univariate, conjunctive, and non-convex) into widespread LLMs like ChatGPT and Llama via system instructions, using global XAI methods to detect them. This way, we found that RuleFit struggles with conjunctive and non-convex biases, while SHAP can approximate conjunctive biases but cannot express them as actionable rules. Hence, we introduce RuleSHAP, a global rule extraction algorithm combining SHAP and RuleFit to detect more non-univariate biases, improving injected bias detection over RuleFit by +94% (MRR@1) on average.
This paper presents an end-to-end suite for multilingual information extraction and processing from image-based documents. The system uses Optical Character Recognition (Tesseract) to extract text in languages such as English, Hindi, and Tamil, and then a pipeline involving large language model APIs (Gemini) for cross-lingual translation, abstractive summarization, and re-translation into a target language. Additional modules add sentiment analysis (TensorFlow), topic classification (Transformers), and date extraction (Regex) for better document comprehension. Made available in an accessible Gradio interface, the current research shows a real-world application of libraries, models, and APIs to close the language gap and enhance access to information in image media across different linguistic environments
Recent studies on large language model (LLM) reasoning capabilities have demonstrated promising improvements in model performance by leveraging a lengthy thinking process and additional computational resources during inference, primarily in tasks involving mathematical reasoning (Muennighoff et al., 2025). However, it remains uncertain if longer reasoning chains inherently enhance factual accuracy, particularly beyond mathematical contexts. In this work, we thoroughly examine LLM reasoning within complex open-domain question-answering (QA) scenarios. We initially distill reasoning traces from advanced, large-scale reasoning models (QwQ-32B and DeepSeek-R1-671B), then fine-tune a variety of models ranging from smaller, instruction-tuned variants to larger architectures based on Qwen2.5. To enrich reasoning traces, we introduce factual information from knowledge graphs in the form of paths into our reasoning traces. Our experimental setup includes four baseline approaches and six different instruction-tuned models evaluated across a benchmark of six datasets, encompassing over 22.6K questions. Overall, we carry out 168 experimental runs and analyze approximately 1.7 million reasoning traces. Our findings indicate that, within a single run, smaller reasoning models achieve noticeable improvements in factual accuracy compared to their original instruction-tuned counterparts. Moreover, our analysis demonstrates that adding test-time compute and token budgets factual accuracy consistently improves by 2-8%, further confirming the effectiveness of test-time scaling for enhancing performance and consequently improving reasoning accuracy in open-domain QA tasks. We release all the experimental artifacts for further research.
Patent claims define the scope of protection and establish the legal boundaries of an invention. Drafting these claims is a complex and time-consuming process that usually requires the expertise of skilled patent attorneys, which can form a large access barrier for many small enterprises. To solve these challenges, researchers have investigated the use of large language models (LLMs) for automating patent claim generation. However, existing studies highlight inconsistencies between automated evaluation metrics and human expert assessments. To bridge this gap, we introduce Patent-CE, the first comprehensive benchmark for evaluating patent claims. Patent-CE includes comparative claim evaluations annotated by patent experts, focusing on five key criteria: feature completeness, conceptual clarity, terminology consistency, logical linkage, and overall quality. Additionally, we propose PatClaimEval, a novel multi-dimensional evaluation method specifically designed for patent claims. Our experiments demonstrate that PatClaimEval achieves the highest correlation with human expert evaluations across all assessment criteria among all tested metrics. This research provides the groundwork for more accurate evaluations of automated patent claim generation systems.