Two questions regarding practitioners' use of patent embeddings arise: (i) Does one fine-tuning recipe suffice for all downstream applications? (ii) Is fine-tuning on one patent landscape sufficient for downstream application on other landscapes? By evaluating 22 pre-trained embedding models (ranging from 22M to 12B parameters) on three tasks -- information retrieval, classification, and clustering -- on 113,148 WIPO patents for assistive technology (46,069 citation queries) and on an external DAPFAM dataset, we find that two results cast doubt on the prevailing wisdom. (i) The optimal fine-tuning recipe depends on the downstream task: cross-sectional alignment (recipe R3) provides the largest improvements to retrieval performance (+7.1% nDCG@10), whereas a combined signal recipe (recipe R4) is better suited to classification (+7.1 F1) and clustering (+10.9 V-measure); a matched data control confirms that differences in training dataset size are not a contributing factor. (ii) Single-landscape fine-tuning hampers cross-landscape information retrieval: fine-tuning on one landscape significantly degrades cross-domain retrieval for 5 of 8 model-recipe combinations on the DAPFAM corpus, with the stronger zero-shot models suffering most. While within-family scaling is consistent (Qwen3 0.6B->4B->8B; Llama-Nemotron 1B->8B), cross-family scaling is erratic; the 12B KaLM-Gemma3 is ranked 8th on TAC retrieval performance, following prefix modification. Title+Abstract+Claims is the ubiquitous best text view, and all models suffer from a 55-65% gap between IN and OUT-of-domain performance which cannot be mitigated by hybrid BM25-dense fusion. Code and evaluation framework are publicly available.
Linear attention replaces the unbounded cache of softmax attention with a fixed-size recurrent state, reducing sequence mixing to linear time and decoding to constant memory. The hard part is not just what to forget, but how to edit this compressed memory without scrambling existing associations. Delta-rule models subtract the current read before writing a new value, and Kimi Delta Attention (KDA) sharpens forgetting with channel-wise decay. But the active edit still uses a single scalar gate to control two different things: how much old content to erase on the key side and how much new content to commit on the value side. We introduce Gated DeltaNet-2, which generalizes both Gated DeltaNet and KDA by inheriting adaptive forgetting and channel-wise decay while addressing their shared limitation, the scalar tie between erasing and writing. Gated Delta Rule-2 separates these roles with a channel-wise erase gate b_t and a channel-wise write gate w_t, reducing to KDA when both gates collapse to the same scalar and to Gated DeltaNet when the decay also collapses. We derive a fast-weight update view, a chunkwise WY algorithm with channel-wise decay absorbed into asymmetric erase factors, and a gate-aware backward pass that preserves efficient parallel training. At 1.3B parameters trained on 100B FineWeb-Edu tokens, Gated DeltaNet-2 achieves the strongest overall results among Mamba-2, Gated DeltaNet, KDA, and Mamba-3 variants across language modeling, commonsense reasoning, and retrieval. Its advantage is most pronounced on long-context RULER needle-in-a-haystack benchmarks, where it improves the evaluated multi-key retrieval setting and remains strong in both recurrent and hybrid settings. Code is available at https://github.com/NVlabs/GatedDeltaNet-2.
We study the Compositional Geometry Routing Problem (CGRP), a unified superclass of traditional routing problems that covers point-only, line-only, area-only, and arbitrary hybrid task geometries, providing a broad abstraction for real-world routing scenarios. Beyond standard point-based routing, CGRP with non-point tasks can be inherently asymmetric, tightly coupled travel routes with the intrinsic path, and enlarges the action space with numerous feasible yet often irrelevant options, thereby posing significant challenges for both representation learning and decision-making. To address these challenges, we propose DiCon, a differential attention-assisted solver with contrastive learning, as a plug-and-play framework that tackles the problem from two complementary angles. First, we introduce a differential attention mechanism that actively suppresses the probability mass on less competitive candidate actions. Second, we design a double-level contrastive learning objective to promote robust global instance representations and regularize geometry-aware task representations. Extensive experiments demonstrate that DiCon achieves strong performance, broad versatility, and superior generalization across diverse CGRP instances with different compositions.
Multi-agent language systems are often built as hand-designed workflows, where agents are assigned semantic roles and communication protocols are specified in advance. We propose NeuroMAS, a framework that treats a multi-agent language system as a trainable and scalable neural-network-like architecture with hidden textual intermediate signals. In NeuroMAS, language-agent nodes are role-free but structure-aware: the topology determines how information can flow, while reward-based training determines how nodes communicate, specialize, and coordinate. This formulation shifts multi-agent design from workflow engineering toward architecture design, where depth, width, connectivity, and growth protocol become scalable sources of capability. We provide a theoretical perspective showing why such modular text-valued computation can be parameter-efficient when tasks admit stable decompositions. Experiments show that NeuroMAS improves over both inference-time collaboration methods and trained baselines, and that its gains are not explained by trainable parameter count alone. We further find that organizational scaling is path-dependent: larger systems can fail when trained from scratch, but become stronger when grown progressively from smaller trained systems. These results suggest that learned multi-agent organization is a promising scaling axis for language-model reasoning and generation.
Generative agents have proven to be powerful assistants in a wide variety of contexts. Given this success, users are now deploying agents with minimal restrictions in open ended, multi-agent environments. Current methods for monitoring the dynamics of open-ended multi-agent systems are limited to aggregate statistics, qualitative inspection, or detecting change between two fixed timepoints. In this paper, we extend the process-theoretic notion of adaptive control charts to multi-agent systems to enable automated monitoring. Using simulation, we demonstrate that adaptive control charts are necessary for monitoring multi-agent systems that can learn from their environment. We further demonstrate, both empirically and theoretically, that adaptive control charts are susceptible to adversarial agents that defect sufficiently slowly. These results suggest a fundamental tradeoff in multi-agent system control: monitors with fixed baselines flag legitimate learning, while monitors with adaptive baselines admit slow adversaries. We conjecture this holds for any monitor.
A growing body of work pursues AI scientists capable of end-to-end autonomous scientific discovery. This position paper argues that although they already function as co-scientists, agentic AI scientists are not built for autonomous scientific discovery. We identify the following challenges in building and deploying autonomous AI scientists: (1) Problem selection is influenced by the McNamara fallacy; (2) Agents are built on large language models (LLMs) whose training corpora omit tacit procedural and failure knowledge of laboratory practice; (3) Preference optimisation during post-training compresses output diversity toward consensus; and (4) Most scientific benchmarks measure single-turn prediction accuracy and lack feedback from physical experiments back to the computational model. These challenges are not just questions of scale and scaffolding; they require revisiting fundamental design choices. To build truly autonomous AI scientists, we recommend scientific simulations as verifiers for training, a persistent, mutable epistemic state that carries beliefs and shifting objectives across an investigation, the establishment of a centralized preregistration repository for all AI-generated hypotheses, and application driven by scientific need rather than tool affordance.
Modern AI is opening the door to collective decision-making in which participants express their views as free-form text rather than voting on a fixed set of candidates. A natural idea is to embed these opinions in a vector space so that the substantial literature on facility location problems and fair clustering can be brought to bear. But standard text embeddings measure semantic similarity, whereas distances in facility location problems and fair clustering require what we call preferential similarity: a participant's agreement with a piece of text should be inversely related to their distance from it. Off-the-shelf embeddings inherit a coarse preference signal through a correlation between semantic and preferential similarity, but fail to capture preferences when the correlation breaks. We formalize this as an invariance problem: text embedding models encode both a preference-relevant signal (stance and values) and semantic nuisance (style and wording), and the two are observationally correlated, so a geometry that relies on nuisance can appear preference-correct even when it is not. We show that synthetic training data designed to break this correlation provably shifts the optimal scorer away from nuisance-dominated cosine and significantly improves preference prediction across 11 online deliberation datasets.
At the heart of a language model agent is an orchestrator program which is responsible for the state transition between consecutive turns. In existing systems, this program is implemented in the harness. This paper introduces self-programmed execution (SPE), in which the model completion is itself the orchestrator program, and the harness evaluates this program but does not impose any fixed orchestration policy. Formally, an SPE state of an agentic machine is one from which a model completion can load any successor state up to an embedding, and an instance of an SPE state is one which executes the model completion as a program. In practice, the program requires a language with unusual features, as its source code doubles as the content of the model's context window. I therefore introduce Spell, a Lisp dialect in which programs can edit and re-evaluate themselves, and effectful expressions like model invocations are structured such that re-evaluating an edited program does not replay its side effects. Experiments with existing models, not trained for SPE or Spell, show that frontier models can learn Spell in-context and accomplish challenging agentic tasks. These results demonstrate how an LM can act as an agent without any fixed orchestration policy, and they raise the question of what self-orchestration strategies might be learned by a model trained for self-programmed execution. Code is available at https://github.com/lukejoconnor/spell.
Most coding-agent benchmarks ask whether generated code behaves correctly. That remains essential, but repository-level engineering is increasingly agent-managed: one agent writes a repository, and later agents inspect, audit, or extend it as working context. In that setting, a generated repository is not only an answer to a task but also a communication artifact for future work. Even when strong agents nearly satisfy the visible behavioral objective, repositories can differ in how clearly they expose the intended behavior and design choices behind that behavior. We introduce Build-and-Find to evaluate whether downstream agents can efficiently recover intended choices from generated repositories. For each task, a builder creates a codebase from a hidden specification; a finder receives only the codebase and a specification-traced question bank. The protocol separates behavioral correctness from artifact-side recovery and reports recovery accuracy, repeatability, implementation coverage, and inspection effort. Evaluation proceeds in two stages: the first uses recovery accuracy and stability as gates to establish that finders can reliably recover the intended choices; the second compares inspection effort by the same finder across artifacts that pass these gates and allow the same intent to be recovered. Lower effort suggests that the artifact makes that intent easier to locate. We find that several frontier models approach saturation in intent recovery on our task pack, yet second-stage inspection effort reveals differences among their generated artifacts, providing a dimension beyond accuracy for comparing artifact legibility. We also audit artifacts to distinguish unimplemented specification claims from finder failures and check whether correct answers are supported by artifact evidence. We release the harness and benchmark data for reproducibility and extension.
Artificial intelligence offers powerful paradigms for scientific discovery, orchestrating autonomous inner search loops over long horizons. However, discovery remains bottlenecked by the friction of the outer human-AI loop: the cycle where domain experts formulate research objectives, diagnose reward hacks, and extract meaning. We present findings from a study with 11 expert mathematicians using AlphaEvolve, an evolutionary coding agent, across more than 2,300 experiments over three months. We identify and characterize intentmaking, the iterative process of discovering, defining, and refining experimental goals through system interaction, as a reciprocal counterpart to sensemaking. We observe that domain experts effectively alternate between steering the search space (intentmaking) and extracting mathematical insights (sensemaking). Our findings suggest an approach to designing AI tools for scientific discovery that moves beyond the existing question--answer model, treating them as collaborative instruments rather than opaque black-box oracles.
Time-series data are inherently multiscale, spanning diverse temporal granularities from coarse trends to fine-scale dynamics. However, existing time-series generative models provide limited control over the temporal granularity of both inputs and outputs, restricting their ability to condition on user-provided coarse sketches and generate samples at a desired target granularity. To address this, we introduce TimeTok, a unified framework for Granularity-Controllable Time-Series Generation (GC-TSG), which generates time series at any target granularity from any coarser input (e.g., rough sketches) or without conditioning. At the core of TimeTok is a hierarchical tokenization strategy that maps time series into an ordered sequence of tokens, from coarse to fine temporal granularity. Our autoregressive generation process operates across these granularity levels, producing token blocks that are decoded back into continuous time series. This design naturally enables GC-TSG within a single framework, where controlling the number of token blocks provides explicit control over output detail. Experiments show that TimeTok excels at GC-TSG tasks while achieving state-of-the-art performance in standard generation. Furthermore, we showcase TimeTok's potential as a foundational tokenizer by training on multiple datasets with heterogeneous temporal granularities, verifying strong transferability that consistently outperforms models trained on individual datasets. To our knowledge, this is the first unified framework that covers the full generative spectrum for time series, offering a valuable foundation for models that benefit from diverse temporal granularities.
We present XekRung, a frontier large language model for cybersecurity, designed to provide comprehensive security capabilities. To achieve this, we develop diverse data synthesis pipelines tailored to the cybersecurity domain, enabling the scalable construction of high-quality training data and providing a strong foundation for cybersecurity knowledge and understanding. Building on this foundation, we establish a complete training pipeline spanning continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL) to further extend the model's capabilities. We further introduce a multi-dimensional evaluation system to guide the iterative improvement of both domain-specific and general-purpose abilities. Extensive experiments demonstrate that XekRung achieves state-of-the-art performance on cybersecurity-specific benchmarks among models of the same scale, while maintaining strong performance on general benchmarks.
We introduce a framework called LAPITHS (Language model Analysis through Paradigm grounded Interpretations of Theses about Human likenesS) and use it to show that several major claims advanced by models such as CENTAUR, proposed as an artificial Unified Model of Cognition, are not theoretically or empirically justified. LAPITHS provides a principled reference point for counteracting the current behaviouristic tendency in AI research to interpret the human level performances of transformer based language models as evidence of human like underlying computation and, by extension, as signs of cognitive abilities. The novelty of LAPITHS lies in making explicit the arguments grounded in two quantitative assessments: (i) the Minimal Cognitive Grid, a theoretically motivated method for estimating the cognitive plausibility of artificial systems, and (ii) a behavioural comparison showing that results similar to those reported for CENTAUR like models can be reproduced by other systems that do not satisfy the structural constraints typically associated with cognitive plausibility, and whose outputs do not provide independent explanatory insight into human cognition.
Skill-distillation pipelines learn reusable rules from LLM agent trajectories, but they lack a key signal: how much each step costs. Without per-step cost, a pipeline cannot distinguish adding a missing step to fix a bug from removing an expensive step that never affected the outcome. We use the cost-attribution gap to ask whether the rule types inside a distilled skill transfer the same way to new tasks. ClawTrace records cost-attributed agent traces and compiles each session into a TraceCard; CostCraft reads TraceCards and writes three kinds of skill patches: preserve, prune, and repair. We find a pattern aggregate metrics hide. On 30 held-out SpreadsheetBench tasks across two seeds, removing prune patches roughly tripled the quality-regression count without lowering median cost. Across the full 84-task SkillsBench transfer, CostCraft saves no aggregate cost. All three quality regressions trace to the preserve lane, and both quality wins trace to the prune lane: prune patches act as quality guardrails while preserve patches drive regressions. We argue that reusable agent skills should be evaluated at the rule-type level, not as monolithic instruction packages. To support this, we release ClawTrace, the TraceCard schema, and the full set of typed skills.
Industrial robots are widely used in manufacturing, yet most manipulation still depends on fixed waypoint scripts that are brittle to environmental changes. Learning-based control offers a more adaptive alternative, but it remains unclear whether such methods, still mostly confined to laboratory demonstrations, can sustain hours of reliable operation, deliver consistent quality, and behave safely around people on a live production line. Here we present Learning-Augmented Robotic Automation, a hybrid system that integrates learned task controllers and a neural 3D safety monitor into conventional industrial workflows. It retains the core industrial automation backbone while introducing learning only where adaptability is needed. We deployed the system on an electric-motor production line to automate deformable cable insertion and soldering under real manufacturing constraints, a step previously performed manually by human workers. With less than 20 minutes of real-world data per task, the system operated continuously for 5 hours 10 minutes, producing 108 motors without physical fencing and achieving a 99.4% pass rate on product-level quality-control tests. It achieved near-human takt time with reduced cycle-time variability while consistently producing high-quality solder joints comparable to those of experienced operators. These results establish a practical pathway for extending industrial automation with learning-based methods.
Music Structure Analysis (MSA) aims to uncover the high-level organization of musical pieces. State-of-the-art methods are often based on supervised deep learning, but these methods are bottlenecked by the need for heavily annotated data and inherent structural ambiguities. In this paper, we propose an unsupervised evaluation of nine open-source, generic pre-trained deep audio models, on MSA. For each model, we extract barwise embeddings and segment them using three unsupervised segmentation algorithms (Foote's checkerboard kernels, spectral clustering, and Correlation Block-Matching (CBM)), focusing exclusively on boundary retrieval. Our results demonstrate that modern, generic deep embeddings generally outperform traditional spectrogram-based baselines, but not systematically. Furthermore, our unsupervised boundary estimation methodology generally yields stronger performance than recent linear probing baselines. Among the evaluated techniques, the CBM algorithm consistently emerges as the most effective downstream segmentation method. Finally, we highlight the artificial inflation of standard evaluation metrics and advocate for the systematic adoption of ``trimming'', or even ``double trimming'' annotations to establish more rigorous MSA evaluation standards.
Language models are increasingly trained to "reason" before answering users' queries, outputting hundreds of intermediate tokens before a final answer. While reasoning is designed with direct model capabilities in mind, we argue that reasoning language models (RLMs) should also be assessed by interaction effects between their reasoning traces and weaker models. In contexts like safety monitoring and distillation, the performance of weak models increasingly depends on legibility--how thoroughly and precisely a reasoning trace externalizes an RLM's actions. Existing efficiency-based metrics for legibility fail to capture "thoroughness", instead focusing on conciseness. Thus, we introduce transfer utility, a method for measuring trace legibility derived from interactions where weak models finish tasks given incomplete traces. Evaluating 84k traces from 12 RLMs across three tasks, we find that reasoning traces generated by the highest-performing and the most efficient models rank the lowest for transfer utility. We also find evidence that transfer utility predicts the ability of weak monitors to verify procedurally dense traces for math or logical problem-solving, though the effect moderates when verifying fact-driven traces. Finally, we show that open-source reward models decouple from transfer utility signals on correctness. Together, these findings surface status quo trade-offs in weak-to-strong legibility, motivating an interaction-first framework for future applications.
Large language models increasingly act as autonomous agents in long-horizon interactions, learning from final outcomes and environmental feedback. However, optimizing from such interaction experience remains challenging, as feedback is heterogeneous and rarely carries explicit reward information. Existing post-training approaches rely on outcome-driven methods, such as reinforcement learning with verifiable rewards (RLVR), which primarily exploit final success signals while leaving interaction experience and feedback underutilized. In sparse-reward, long-horizon tasks, this often results in distribution sharpening: the policy reinforces a narrow set of already-successful behaviors, without substantially improving the feedback-grounded agency needed for broader problem-solving capacity (e.g., Pass@k). We propose Counterfactual Distillation, a framework that converts exploration-derived experience into trainable policy supervision. Specifically, we organize exploration as a tree-structured search process, where the agent reflects on past decision points and makes new experience-guided decisions to form alternative branches. These corrections are then distilled as counterfactual targets: we train the model to predict the revised action from the original history without explicitly providing the experience, thereby introducing out-of-distribution behavior patterns and expanding exploration capacity. Across diverse interactive coding and agentic tasks, our method outperforms outcome-driven baselines such as GRPO and experience-based methods such as Early Experience, with gains of up to 14 percentage points over the base model on Pass@128. When interleaved with reinforcement learning updates, it further raises the performance ceiling, improving Pass@1 on Sokoban by 8 percentage points over extended GRPO. Our code is available at https://github.com/motigrez/Counterfactual-Distillation.
Preference-based reinforcement learning (PbRL) is the dominant framework for aligning AI systems to human preferences. However, evaluation protocols for such data were designed for text and have not been validated for speech. We present the first intraclass correlation coefficient (ICC)-based, controlled cross-modal study of human and synthetic preference annotations, comparing text and audio evaluations of identical semantic content across 100 prompts. We show that achieving \textit{good} (audio) to \textit{near good} (text) agreement within either modality (ICC(2,$k$) $\approx$ .81 audio, .74 text) requires $\sim$9 raters and modalities show marked differences in how people report preferences: audio raters exhibit narrower decision thresholds, reduced length bias, and more user-oriented evaluation criteria, with near-chance cross-modality agreement. We demonstrate that synthetic ratings can be used to effectively predict inter-rater agreement, thus serving as an early signal for stimulus selection and proxy for human annotations. Together, these findings argue that evaluation protocols for audio preference data require modality-specific design rather than direct adaptation from text. Code, data, and the audio corpus are available at: https://huggingface.co/datasets/sciminds/same-words-different-judgements
This study aimed to comprehend how user domain knowledge and artificial intelligence (AI) literacy impact the effective use of human-AI interactive building energy management system (BEMS). While prior studies have investigated the potential of integrating large language models (LLMs) into BEMS or building energy modeling, very few studies have examined how user interact with such systems. We conducted a systematic role-playing experiment, where 85 human subjects interacted with an advanced generative pre-trained transformer (OpenAI GPT-4o). Participants were tasked with identifying the top five behavioral changes that could reduce home energy use with the GPT model that functioned as an LLM-integrated BEMS. Then, the collected prompt-response data and participant conclusions were analyzed using an analytical framework that hierarchically assessed and scored human-AI interactions and their home energy analysis approaches. Also, participants were classified into four groups based on their self-evaluated domain knowledge of building energy use and AI literacy, and Kruskal-Wallis H tests with post-hoc pairwise comparisons were conducted across 20 quantifiable metrics. Key takeaways include: most participants employed concise prompts (median: 16.2 words) and relied heavily on GPT's analytical capabilities; and notably, only 1 of 20 metrics, appliance identification rate, showed statistically significant group differences (p=0.037), driven by AI literacy rather than domain knowledge, suggesting an equalizing effect of LLMs across expertise levels. This study provides foundational insights into human-AI collaboration dynamics and promising development directions in the context of LLM-integrated BEMS and contributes to realizing human-centric LLM-integrated energy systems.