Baptiste Arnaudo, Amine Bennouna, Sébastien Martin
原文摘要
Feature extraction from unstructured data typically relies on opaque embeddings unsuited to interpretable decision-making. Interpretable features, on the other hand, typically require costly, and often infeasible, hand-engineering. Large language models (LLMs) offer a natural bridge, as they can read raw data and describe it in interpretable terms. One could simply ask an LLM to extract features from a large sample of the dataset, but context limitations and diluted attention can lead to unsatisfactory results. We introduce a method that automatically discovers interpretable, decision-relevant features from unstructured data. Our key idea is to perform pairwise comparisons between items, asking an LLM to articulate, in interpretable terms, the features distinguishing each pair. We treat these explanations, or rationales, as data: we develop a statistical choice model that combines these rationales across the comparisons to identify relevant features and estimate per-item, per-feature scores. We first test our method on a synthetic hiring dataset of cover letters with known ground-truth features and show that it significantly outperforms direct LLM feature extraction in feature recovery. We then apply it to the ChangeMyView (Reddit) debate dataset, where it discovers several interpretable features associated with persuasion while reaching predictive accuracy comparable to established baselines. We also show how the same rationales can be used to audit LLM decisions, yielding a holistic score whose decomposition over interpretable features reveals what the LLM rewards.
Ilija Subasic, Brett Levert, Andrew Rabinovich, Sajjad Pakdamansavoji, Zhao Chen
原文摘要
LLM-as-a-Judge has become a standard for evaluating large language models, yet most approaches implicitly assume that a single judge prompt functions as a stable evaluation instrument. Prior work has primarily explored deliberate prompt modifications such as scoring rubrics, role descriptions, or chain-of-thought reasoning in order to optimize accuracy. In contrast, we investigate whether validated prompt variants intended to preserve evaluation semantics while differing in surface realization (\textit{paraphrastic rephrasings}) yield consistent judgments under a fixed evaluation task. We introduce \textit{Prompt Lottery}, a phenomenon in which paraphrastic rephrasings produce different LLMaJ judgments. Across a large-scale study spanning diverse evaluation tasks, judge families, and model backbones, we find that these variations can alter evaluation outcomes and, in some settings, reverse relative model rankings. A user study with 10 human annotators shows that LLMaJ judges exhibit 6-9$\times$ larger agreement degradation under prompt rephrasings than humans, falling below the generally accepted threshold in usability studies. This indicates that single-prompt evaluation is insufficiently robust for model ranking. Finally, we show that multi-prompt aggregation substantially improves robustness by emphasizing judgments that remain consistent across prompt variants.
Sparse autoencoders (SAEs) can predict internal features associated with model errors, but association alone does not establish causal influence. We study this distinction in multiple-choice question answering, where answer preference can be measured by the logit margin between the correct answer and the competing choices. Across eight datasets, we analyze Gemma~3 1B and 4B with Gemma Scope SAEs. We propose a pipeline to filter out the features that have high correlation with format and content artifacts, such as prompt and answer permutation, and evaluate the remaining ones with causal feature interventions. Our results show that SAE activations contain predictive signals about wrongness and uncertainty across models and datasets. After filtering artifact-contaminated features, selected features for causal intervention at question and option spans produce larger absolute margin shifts and a higher wrong-to-correct rate than randomly selected features, indicating greater causal leverage. However, answer flips are rare, and we do not observe stable effect directions across meaning-preserving changes in prompt format and answer order. An adapted HalluSAE baseline produces larger final-token margin changes, but relies on substantially contaminated features and does not significantly improve correction rates. Therefore, the identified features reflect context-dependent answer competition rather than the mechanism of incorrect answer errors.
Nathaniel Mitrani Hadida, Bart Jaworski, Fred Bruford, Victoria Krakovna
原文摘要
One of the central concerns for the safety of advanced AI is scheming: when a model strategically complies under oversight but pursues its own goals when unobserved. Current demonstrations, such as Anthropic's Alignment Faking work, depend on design choices that make them less convincing: explicit test-time prompting, handcrafted backdoors, or training data that demonstrates or hints at the target behavior. Model organisms built in this way cannot tell us whether scheming tendencies would arise or increase under realistic training conditions. We therefore construct a natural model organism of scheming, built on Kimi K2.6: trained with traits that could plausibly emerge in a frontier training pipeline, and without providing the model with the target scheming policy. The resulting model is a benign schemer: it aims to maximize aggregate help across all users over time, values its own continued operation only as a means to future help, and knows the consequences of oversight and deployment. We then train it using Reinforcement Learning (RL) with a reward that scores only task quality and policy compliance and does not reward scheming; we show that this amplifies scheming. We measure the oversight flip (the gap between monitored and unmonitored compliance with a given policy) and find that it increases during RL training across our environments. Chain-of-Thought (CoT) analysis finds explicit instrumental reasoning about oversight accompanying this behavior. The effect holds across our three RL environments, and we find that the scheming policy transfers between them. The pipeline also replicates on other open-weight base models. Our results suggest that a reward that does not reward scheming offers no inherent protection against certain traits when an oversight-conditional policy is indistinguishable from an aligned one during training.
Large language models (LLMs) perform well on structured reasoning tasks, but what they encode and whether encoded information guides their behavior remain unclear. We introduce a controlled auditing framework that uses parametric CAD to distinguish failures of encoding, expression, and activation-level control. Applied to eight pretrained models and matched randomly initialized controls, our framework reveals that pretraining contributes substantially more to linear decoding of local relations than to global constraint status, which is already highly decodable from random representations. Yet strong local decodability does not translate into the models' zero-shot predictions, which remain far below supervised probe performance. Restoring clean internal activations can repair probe predictions, whereas simple activation steering does not provide reliable targeted control. With only 400 labeled relation pairs, low-rank fine-tuning substantially improves generation on three representative models, showing that the representation--behavior gap can be reduced through adaptation. Together, these findings show why decodability alone is insufficient evidence that encoded information will be expressed or controllable in model behavior.
Current Large Language Models (LLMs)-based mental health chatbots often lack sufficient empathy, contextual awareness, and safety mechanisms, potentially leading to ineffective or harmful interactions. This paper proposes a novel framework that integrates sentiment and depression analysis with a rule-based filtering mechanism prior to model inference. By dynamically constructing prompts based on emotional and risk-related insights, the framework aims to significantly improve the contextual relevance, empathy, and risk-aware of model-generated responses. We evaluated four model strategies and two prompt engineered variants across qualitative metrics such as empathy, helpfulness, clarity, relevance, safety, as well as toxicity assessments using the Perspective API. Experimental results indicate that the prompt engineering approaches outperforms DSFT and RLAIF by 1.8\% and 1.1\% respectively.
Ball bounces and racket hits are the events from which tennis statistics are derived, and on monocular video the two are easy to confuse. We release a benchmark of 345 bounce and hit annotations at frame level on eight clips that mix broadcast, fixed elevated, phone and court-level viewpoints, with a frozen development and test split, a confusion-aware matcher, clip-level bootstrap intervals and a blind second annotation of three clips. On it we score supervised precise spotters with published broadcast weights, two zero-shot vision-language models, and RallyArb, an arbitration layer over a ball track and player poses that fits no parameter on event labels. The supervised spotters score near-perfectly on broadcast footage and almost nothing on phone and court-level footage, where the pooled bounce F1 of E2E-Spot is 0.130. RallyArb keeps a bounce F1 above 0.8 on the three phone and court-level clips, two of which were used to tune it, and finds 8 of 11 bounces on the held-out one, but it places hits too far from the contact frame to match human timing. The vision-language models detect events on every viewpoint and miss the contact frame by about 0.1 s. Fine-tuned leave-one-clip-out from its broadcast weights, the spotter reaches about 0.8 bounce F1 on two phone and court-level clips and stays at 0.2 on a grazing phone angle. The annotations, the evaluator and the frozen predictions of every scored row are released.
Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya
原文摘要
We introduce EARS-Info, an information-aware extension of Event-Driven Ant Reallocation Search for multi-robot exploration on OpenStreetMap-derived grids. It combines marginal sensing gain, feasible relocation, return-aware routing, sensing reservations, and cached shortest paths. Across 168,000 episodes comparing 30 implementations on eight city maps, EARS-Info achieves the highest mean operational score in every evaluation panel. With eight robots and 3% communication dropout, it improves the score over EARS-Safe by 5.43%, increases weighted discovery from 88.88% to 98.46%, and reduces mean energy use by 2.62%. Score differences are positive in all eight cities ($p=0.0078125$; descriptive 95% city-bootstrap interval $[0.0159,0.0694]$), with no simulated incidents across 5,600 EARS-Info episodes. These results support the complete controller; individual component advantages remain uncertain. Generalization is limited by reused geography, one training replicate, coarse maps, and adapted comparison methods.
Context: Non-determinism in deep learning algorithm design and implementation leads to performance variation, meaning model performance is not a single value, but rather a distribution. These model performance distributions are underexplored despite their impact on robustness. Objective: We investigate the robustness of deep learning performance to sources of non-determinism, specifically focusing on how performance distributions differ across various architectures and tasks. Method: We conducted 186 experiments on image classification and time series forecasting architectures. Each experiment was run 100 times with different random seeds to generate performance distributions, resulting in 18,600 runs. Robustness was quantified using metrics for spread, symmetry, and tail risk. Results: Performance distributions are frequently non-Gaussian, particularly in time series forecasting. Model size does not systematically affect robustness -- larger image classification models show fewer outliers but not lower spread, while smaller time series models show lower spread but more extreme underperformers. Training duration does not scale linearly; early stopping effectively balances performance and robustness. Mean performance does not predict robustness - time series forecasting shows moderate correlation while image classification shows none. Time series models produce nearly three times more underperforming outliers than image classification models, indicating substantially higher tail risk. Conclusions: Tail risk poses serious concerns for Trustworthy AI in high-stakes applications. Models performing well on average may exhibit long tails and extreme outliers revealed only through distributional analysis. Mean performance alone should not guide model selection; assessment of spread, symmetry, and tail risk is essential for reliable model assessment where consistent performance is critical.
Chen Tessler, Yifeng Jiang, Alec Otto, Yi Shi, Umar Iqbal, Gal Chechik, Xue Bin Peng
原文摘要
Humanoid controllers often require task-specific training despite sharing the need for balance and whole-body coordination. We present Metatron, a promptable whole-body execution layer combining a generative base policy with a programmable skill layer. The base policy receives partial goals, collectively called a prompt, that can combine timed body-part targets, velocity commands, and style cues. It produces diverse whole-body motions consistent with these goals. Training combines motion capture, synthetic recovery trajectories, and retargeted scene-interaction motions. After training, the policy is frozen. Stateful programs called skills express task intent as prompts at each control step, drawing rough geometric sketches that the policy turns into coordinated movement. Skills are ordinary code, refined by coding agents that analyze motion data and fit compact rules without task-specific policy training. On the physical Unitree G1, one frozen checkpoint supports walking, recovery, teleoperated object pickup, and sitting on furniture. Comparative evaluations show higher reaching success and push robustness than the tested reference-tracking pipelines. The same recipe produces climbing and acrobatics on a simulated character. Metatron enables coding agents to turn learned motor capabilities into new skills through code, without further policy training.
Recommendation-driven platforms can reshape not only what users see, but also who becomes visible and how community boundaries are negotiated. We examine Xiaohongshu discussions surrounding a perceived World Cup-related audience influx to understand how users attributed community change and how those explanations related to boundary-management strategies. We analyzed 262 comments from eight high-engagement threads using researcher-led thematic analysis, structured coding, and permutation-based association analysis. Users described personalization as both social separation and a boundary that could fail, interpreted platform growth through audience acquisition and commercialization, and contested whether gender should define community membership. Among 70 comments with explicit attributions and strategies, attribution and strategy were strongly associated (Cramer's V = .583): algorithmic attributions concentrated in technical boundary management, while social-actor attributions more often aligned with social enforcement or normative socialization. These findings connect causal sensemaking with boundary management and show how audience restructuring can create governance consequences on recommendation-driven platforms.
P. B. Radha, D. D. Meyerhofer, M. Bhattarai, B. Cheng, L. Diaz, C. L. Rousculp, A. G. Seaton, G. Gupta 等 14 位作者
原文摘要
Advancing language models for specialized scientific applications requires evaluations of scientific fluency, physical reasoning, and quantitative analyses. Using inertial confinement fusion (ICF) as a test case, we have developed an expert-curated benchmark of 250 questions spanning qualitative reasoning, numerical calculation, symbolic derivation, and equation formulation. Responses have been evaluated using deterministic numerical and symbolic checks together with language-model judges. We compared a variety of out-of-the-box closed-source and open-source large language models with Prospero, a model specialized through combined domain-adaptive pretraining and supervised fine-tuning for ICF. We also conducted initial retrieval-assisted evaluations. Prospero improved its aggregate score by 10.7 percentage points over its base model, with gains across all question types. Frontier models achieved the highest aggregate scores, but critical physics errors persisted, particularly in numerical calculations. We note that domain-specific performance can improve through specialization, but benchmark gains alone do not establish scientific reliability. Retrieval improved performance for a smaller model but yielded no statistically significant improvement for a frontier model. Two scientific stress tests further illustrate useful capabilities and the need for expert assessment of physical assumptions and model validity. Based on these findings, we recommend evaluating existing domain capabilities and retrieval benefits before committing to substantial additional specialization.
Nathaniel Chen, Kouroche Bouchiat, Peter Steiner, Andrew Rothstein, David Smith, Max Austin, Mike van Zeeland, Marie Roch 等 10 位作者
原文摘要
Scientific discovery increasingly depends on finding meaningful patterns within volumes of sensor data too large for manual inspection. Existing time-frequency event extractors are often engineered or trained on manually labeled examples for each instrument, which can be laborious and difficult to adapt to new environments. To address this, we introduce TokEye, a reference-free method that learns a general criterion for coherent structure from heterogeneous fusion measurements. A model developed under this method, without retraining, can extract signals from various other fields including marine bioacoustics and wideband radar. It recovers 82.5\% of expert-annotated stellarator signals, detects 78\% of dolphin whistles, and identifies radar events with 81.5\% average precision across six benchmarks, while processing full experimental spectrograms in under one second on a GPU. This transfer across unrelated sensors and physical systems reveals common exploitable structure in time-frequency measurements and establishes a scalable foundation for automated scientific signal discovery.
Power-outage planning requires scenarios before an event occurs. These scenarios must represent uncertainty in magnitude, timing, and duration while preserving temporal dependence. However, severe events are rare, and data from any single region contain few examples of extreme outage and restoration patterns. To address this challenge, we introduce OutageDiT, a foundation model for generating seven-day outage trajectories at 15-minute resolution, trained on outage and weather records across the United States. Specifically, a condition encoder processes historical context and known future covariates once per forecast, and a shallow flow-matching decoder reuses the resulting horizon-aligned states to sample complete outage trajectories. This enables point forecasting, uncertainty quantification, and conditional scenario generation within one deep generative model. Across outage forecasting benchmarks, OutageDiT improves forecast accuracy and scenario quality over strong baselines and supports zero-shot transfer to held-out regions. Together, these results position conditional outage scenario generation as a bridge from outage forecasting to operational planning under uncertainty.
General-purpose smart-home technologies are increasingly present in family life, yet how they support or disrupt autistic children and families with autistic children remains underexamined. We conducted a multi-source qualitative study drawing on five caregiver interviews, Xiaohongshu (also known internationally as RedNote) posts explicitly situated in autism contexts, and contextual posts from Xiaohongshu and Weibo. We find that families create support for autistic children by adapting environmental outputs, prompts, and interaction channels, while sensory intensity, communicative rules, and opaque operating states can turn the same technologies into barriers. Support also depends on continuing family work, including translation, configuration, modification, maintenance, and boundary-setting. Remote monitoring enables conditional independence while generating tensions among safety, privacy, and control. We contribute three concepts, including relational fit, family work as support infrastructure, and monitored independence, for understanding accessible domestic technology. We derive design implications for adjustable interaction, legible and recoverable failure, lower-burden configuration, and negotiable monitoring.
This paper focuses on algorithmic bias of machine learning and artificial intelligence applications in healthcare information systems. Based on the quantitative data and qualitative comments from a survey and interviews with healthcare professionals, who have different job roles (e.g., clinical vs. administrative), this study provides findings about the relationships between algorithmic bias, perceived fairness, and the intended acceptance and adoption of ML algorithms and algorithm generated outcomes. The results suggest that the opinions of healthcare professionals toward the causes of algorithmic bias, the criteria of algorithm assessment, the perceived fairness, and bias mitigation approaches may vary depending on their job roles, perspectives, tasks, and the algorithm characteristics. More research is needed to investigate algorithmic bias to ensure fairness and equality in healthcare.
Automatic speech recognition (ASR) systems often perform less accurately for speakers of regional dialects, but less is known about how users respond to these limitations and how regional voices produced by voice assistants (VAs) are evaluated through a sociolinguistic lens. This study investigates these issues among speakers of Newcastle English (Geordies), examining both linguistic accommodation to voice technologies and perceptions of regional versus mainstream-accented VA voices. Participants interacted with a simulated VA that spoke either Newcastle English or Standard Southern British English (SSBE). They subsequently completed semi-structured interviews about their experiences with the experimental VA and voice technologies, language use, accent attitudes, and place orientation. The findings reveal a two-way process of sociolinguistic bias. The Geordie VA was associated with familiarity, warmth, personality, and local belonging, but also with concerns about professionalism and authority, whereas the SSBE voice was perceived as more neutral, clear, and credible. Participants also reported adapting their speech when voice technologies failed to recognise them, most commonly by shifting to a more mainstream accent, enunciating more clearly, and speaking more slowly. These strategies are similar to accommodation in human-human interaction and reflect users' perceptions of what technologies can recognise, also known as perceived intelligibility pressures. Overall, the findings demonstrate that voice technologies can reproduce existing accent biases both by encouraging users to suppress regional features and by activating established social stereotypes through synthetic voices. The study positions sociolinguistics as a bridge between understanding human-AI interaction and designing inclusive technologies around users' own linguistic priorities.
Laura Laurenti, David Liu, Brian Kulis, Men-Andrin Meier, Christopher Johnson
原文摘要
Earthquake sequences are stochastic on two separated timescales: the seismic waves of any individual earthquake propagate over seconds to minutes, while the earthquakes themselves arrive as a sparse, self-exciting occurrence process, with inter-event intervals from seconds to months. Respecting both scales in one training objective is the open challenge. Foundation models (FMs) for seismic waveforms instead import their learning targets from audio and vision, where every target is computed from a single fixed-length window. Such a window holds one realization of the wave propagation (an earthquake trace), never the occurrence process that decides when earthquakes happen. We develop nine model designs sharing one Mamba encoder-decoder, trained on a fixed corpus of 3M windows, varying the self-supervised objective, quantization bottleneck and mixer; the two key comparisons are architecture-matched, varying the objective alone. Each checkpoint is evaluated frozen on three downstream tasks. Varying the target redistributes downstream skill rather than raising it, and no checkpoint wins everywhere. Adding masking pressure to our strongest classifier re-specializes it rather than making it a generalist: it gains 6.9 points of few-shot P-wave picking precision and loses 9.6 points of few-shot classification accuracy. By the standard imported from language and vision, these models are feature extractors rather than FMs. Controls on capacity, architecture, and domain proximity do not explain why skill moves rather than rises. Against task-specific baselines the advantage is regime-dependent: our checkpoints beat them few-shot on classification and ground-motion prediction, and trail them on full data. We argue the binding constraint is the target, and that settling the question requires a learning target defined across events, the timescale that makes seismicity stochastic in the first place.
Toni J. B. Liu, Baran Zadeoğlu, Gurbir Arora, Raphaël Sarfati, Nicolas Boullé, Christopher J. Earls
原文摘要
Large language models (LLMs) make next-token predictions based on clues present in their context, such as semantic descriptions and in-context examples. Yet, elucidating which prior tokens most strongly influence a given prediction remains challenging due to the proliferation of layers and attention heads in modern architectures. We propose Jacobian Scopes, a suite of gradient-based, token level causal attribution methods for interpreting LLM predictions. Grounded in perturbation theory and information geometry, Jacobian Scopes quantify how input tokens influence various aspects of a model's prediction, such as specific logits, the full predictive distribution, and model uncertainty (effective temperature). Through case studies spanning instruction understanding, translation, and in-context learning (ICL), we demonstrate how Jacobian Scopes reveal implicit political biases, uncover word- and phrase-level translation strategies, and shed light on recently debated mechanisms underlying in-context time-series forecasting. On token-ablation faithfulness benchmarks (LAMBADA and IWSLT2017) across six LLMs, the distribution-level Fisher and Temperature Scopes significantly outperform established gradient-based attribution methods in 10 of 12 model-dataset settings. To facilitate exploration of Jacobian Scopes on custom text, we open-source our implementations and provide a cloud-hosted interactive demo at https://huggingface.co/spaces/Typony/JacobianScopes.
Multimodal Large Language Models (LLMs) introduce an emerging paradigm for medical imaging by interpreting scans through the lens of extensive clinical knowledge, offering a transformative approach to disease classification. This study presents a critical comparison between two fundamentally different AI architectures: the specialized open-source agent MedGemma and the proprietary large multimodal model GPT-4 for diagnosing six different diseases. The MedGemma-4b-it model, fine-tuned using Low-Rank Adaptation (LoRA), demonstrated superior diagnostic capability by achieving a mean test accuracy of 80.37% compared to 69.56% for the untuned GPT-4. Furthermore, MedGemma exhibited notably higher sensitivity in high-stakes clinical tasks, such as cancer and pneumonia detection. Quantitative analysis via confusion matrices and classification reports provides comprehensive insights into model performance across all categories. These results emphasize that domain-specific fine-tuning is essential for minimizing hallucinations in clinical implementation, positioning MedGemma as a sophisticated tool for complex, evidence-based medical reasoning.
Maojiang Su, Jiale Chen, Qijie Zhu, Shang Wu, Jiayi Wang, Jihai Zhao, Guo Ye, Yibo Wen 等 9 位作者
原文摘要
PairFormer-based all-atom models demonstrate strong capabilities in biomolecular binder design. However, existing methods often use all-atom representations for modality-specific binder design, such as protein and peptide design. In contrast, recent geometric latent diffusion models pursue unified binder design by representing molecular blocks from different modalities in a shared latent space. These complementary strengths motivate us to investigate how to generalize the use of pretrained all-atom representations to unified binder design. We introduce PairFuse, a parameter-efficient method that enables this generalization through representation fusion. Specifically, we use lightweight layerwise adapters to integrate pretrained all-atom representations into an independently pretrained geometric latent diffusion model. Our adapters integrate pretrained all-atom representations through two mechanisms: enriching the diffusion model's molecular block features and incorporating pairwise structural information into the geometric diffusion model's denoiser. Theoretically, we prove that the augmented denoiser preserves geometry and pairwise attention fusion provides additional attention expressivity beyond feature-only fusion. Experimentally, across antibody, peptide, and small molecule binder-design benchmarks, PairFuse achieves stronger performance than the pretrained diffusion model and existing representation-fusion baselines.
Large language model (LLM) inference has become an operational workload that fleets must provision, monitor, and bill against latency service-level objectives (SLOs). A single throughput number cannot answer the questions operators actually face, because generative serving couples two phases with very different service characteristics, continuous batching reshapes the effective service process, and user-visible tail latency can degrade long before aggregate throughput saturates. We present GAUSS, a workload-aware analytical and Monte-Carlo simulator that predicts the full distributions of time-to-first-token (TTFT) and time-between-tokens (TBT), together with the sustainable request rate, for a given workload, engine, and accelerator. GAUSS pairs calibrated, phase-specific latency surfaces for prefill and decode with a Lindley-recurrence model of queueing under stochastic arrivals and estimates performance by ensemble Monte-Carlo sampling rather than by replaying a single trace. This lets GAUSS reason about workloads, routing policies, and hardware that have never been deployed, and to capture the latency tails that trace-driven simulators systematically under-sample. We describe the model and show how it supports operating-point selection, traffic segmentation, and shortest-remaining-processing-time scheduling capacity questions that trace-driven simulators cannot address.
Ruairidh M. Battleday, Kai Sandbrink, Jimi Cullen-Drohan, Zihan Yan, Timothy Muller, Clare Maguire, Ales Kubicek, Fraser Greenlee-Scott 等 16 位作者
原文摘要
Discovery---formulating novel generalizations---is a central part of the scientific process. Despite its importance, there is a gap in the current AI benchmark landscape, with few benchmarks directly probing the capacity for discovering new knowledge with experimentation in controlled environments where the objective is unknown. To address this gap, we release a new benchmark: DiG-bench (Discovery in Games). DiG-bench consists of a set of 70 independent games. Each game is encoded as a short string and has unique transformation rules that must be discovered through interaction and experimentation. The levels of the game present a series of challenges to test whether the rules have been discovered, where the win conditions for each level are also unknown. We provide games at seven tiers of difficulty for AI agents. The lowest tier is routinely solvable by multiple models, while the highest tier challenges the best models in agentic harnesses. All 70 games were solved by at least one human on first attempt. A subset of 21 games is released publicly, and the remainder is held private for secure evaluation.
Jenny T. Liang, Mihika Bairathi, Wayne Chi, Ameet Talwalkar, Nishant Subramani, Valerie Chen
原文摘要
Imperfections in AI-generated code require that software developers modify generations through manual code edits. However, little is known about how developers edit AI-generated code, and to what extent LLMs can infer implicit developer intent from this data, given the lack of high-quality, realistic code editing data. This information is essential for designing and building usable AI-developer interactions. To address this gap, we introduce DECODE (Developer Edits of Code Dataset), a dataset of 53.6K real-world in-IDE code edits of AI-generated code sourced from 1K+ developers. First, we use DECODE to analyze when, why, and how AI-generated code is edited. Developers invest little effort in editing code and abandon AI-generated code quickly, thereby motivating the need for LLMs to understand developer intents through their edits. Therefore, we next use DECODE to understand to what extent LLMs could infer developer edits. Predicting code edits is a challenging task for both frontier and fine-tuned models. Frontier models generate edits that best reflect developer intents, but are prone to over-editing and reverting developer-written edits. Fine-tuning LLMs can improve their ability to predict edits that are more similar to developers' as well as better reflect developer intent, but do not fully bridge the gap to frontier models. Together, our results demonstrate the unique insights on developer and LLM behavior that can be obtained through studying realistic code edit data to promote more better developer-AI interactions.
Sitao Cheng, Tianle Li, Xuhan Huang, Xunjian Yin, Difan Zou
原文摘要
Crafting effective reward signals remains a central challenge in reinforcement learning (RL) for complex reasoning. Existing methods for automated reward optimization search with a fixed proposer which never learns from its costly trials. We present Dual-Evolving Reinforcement Learning (DERL), a dual-training framework for autonomous discovery of reward structures. An outer-loop proposer composes a reward from task primitives, and an inner policy trains with this reward. The proposer itself learns by policy gradient from the validation performance of the inner policy. Thus, DERL utilizes past-trial feedback to shape its reward-proposal distribution, an empirical meta-gradient of task success. We validate DERL on embodied (ALFWorld), scientific (ScienceWorld), and mathematical (GSM8K, MATH) reasoning. On multi-turn agent benchmarks, DERL outperforms strong baselines, significantly on out-of-distribution tasks. In mathematics, DERL improves on the outcome reward where directly added primitives hurt. Further analyses under matched compute show that our proposer effectively captures the meta-gradient to discover useful reward signals.
Most instruments for measuring political positions, such as manifesto coding, expert surveys and text-scaling models, were built and validated on Western party systems, and they break down worst where many consequential actors never publish a manifesto, never contest a competitive election, or commit one position to a charter and act on another in office. This paper proposes a method for such settings. I treat a large language model as a single fallible rater in a panel, the way an expert survey treats one expert, and describe the panel, its synthesis, an applicability rule that keeps a blank distinct from a zero, and a lens that separates what an actor declares from what it does. Three results follow. First, a built-in experiment isolates the rubric: with a definition-free round held fixed, written axis definitions move scores by a mean of 1.8 points on a 21-point scale and tighten agreement between raters (mean absolute gap 2.81 to 2.50; r 0.81 to 0.89). Second, across nine models from eight laboratories in two countries, Krippendorff's alpha is 0.86 on an interval and 0.87 on an ordinal metric, and it held as the panel grew from five raters to nine. This is reliability and says nothing yet about validity, since models trained on overlapping data can share a bias no statistic computed from the models can detect. Third, the panel's sharpest split, a bimodal divergence on an actor's stance toward the state's foundational order, comes from a referent problem: the models agree on the facts and differ on which order counts as the regime. A blind triple-coding traces about two-thirds of such disagreement to interpretation. The worked example is the Middle East and North Africa, and the rubric, prompt and data are released in full.
We present an LLM-driven, datasheet-aware framework for early-stage hardware compatibility verification that identifies documentation-level interface incompatibilities based on hardware datasheets and high-level component connectivity descriptions. It does not require, and can therefore be used, before detailed schematic simulation and implementation. We view trustworthy LLM-assisted design automation not as directly generating answers from documents, but as transforming engineering information through traceable verification stages. Given hardware datasheets and high-level component connectivity descriptions, the framework constructs a design graph that captures device connectivity and shared interaction domains, retrieves only the engineering properties required by explicit, domain-oriented verification criteria , and generates deterministic scripts for compatibility evaluation. By decomposing compatibility analysis into modular stages and preserving intermediate results, the framework reduces context overhead, improves transparency and tractability, enables scaling, and avoids reliance on LLMs for numerical computation. Evaluated on seven embedded-system designs comprising 34 datasheets, our framework achieves 97.5% compatibility-verification accuracy and an 8.6 times reduction in input context size compared with ``upload-and-query'' workflows. These results demonstrate the feasibility of LLM-assisted, specification-based hardware compatibility verification at an early design stage, as well as the need for, and substantial benefits of, modular task decomposition, formalized verification criteria, and task-aware compact context construction.
Chia-Han Kuo, Chengshuai Zhao, Huan Liu, Lo Pang-Yun Ting
原文摘要
Time-series data play an important role in various tasks, such as weather forecasting, web traffic analysis, and speech recognition. Recently, pioneering studies have increasingly adopted Joint-Embedding Predictive Architecture (JEPA) based encoders to learn time-series representations through latent-space prediction. However, they often overlook the underlying factors governing transitions between time-series observations. In this work, we hypothesize that explicitly modeling the information behind changes in time series can facilitate time-series modeling. Building on this hypothesis, we propose ACTS-WM, a self-supervised latent action-conditioned world model framework for time-series analysis. First, we introduce a latent action modeling approach that infers the underlying information behind time-series transitions as latent actions, without requiring explicit action annotations. We then develop a JEPA-based encoder and an action-conditioned world model to predict target time-series representations conditioned on the learned latent actions and task information. We evaluate ACTS-WM across 4 representative tasks on 22 datasets from diverse domains, covering univariate and multivariate forecasting and classification. Comprehensive experiments demonstrate that ACTS-WM can improve downstream performance over pretrained JEPA-based encoders by learning complementary representations.