We build on EndoWake, an endogenous wake model that links the wind speed at every grid cell to its upwind neighbors by linear constraints, so that wake field, siting decisions and project constraints share one mixed-integer model. Its four parameters are calibrated once on the single wake of a Reynolds-averaged Navier-Stokes (RANS) simulation, and a graphics processing unit (GPU) scores over a million layouts per second per wind direction. At Lillgrund, we compare searches guided by EndoWake and by PyWake under a RANS judge that no search calls: the fidelity of a wake model is not the quality of the layouts optimized under it. Two diagnostics separate a directional quadrature error between the sampled wind directions from a spatially localized surrogate bias off each wake edge, not reported before to our knowledge; at random directions, five paired starts detect no difference between the pipelines. The main result puts the speed to use: EndoWake ranks the moves of a local search on the GPU, and PyWake confirms every accepted move. On twenty unseen starts, with one GPU and thirty CPU processes, screen-and-confirm reaches similar mean gains to a search on PyWake alone, 7.9 times faster at one wind speed and 14.9 times faster on the annual energy (3.0 and 4.7 with a proof of local optimality charged to the timed search); on a second site it keeps its speed, and its local optima fall slightly short of those of PyWake alone. Code and data are deposited on Zenodo.
The records that could train the most useful models, such as a hospital's patient files, are often the ones an institution cannot share. We propose CoRTeC, a method that releases synthetic versions of such records under differential privacy. Differential privacy places a proven limit on what the synthetic data can reveal about any person, and in particular on membership inference. CoRTeC spends the privacy budget once, on a small file of noisy counts designed for training a classifier. For each group of records, the file gives the group's size, its balance of outcomes, a histogram of every column for each outcome and the outcome rate in each of several non-overlapping cells. A reviewer can check every number before the file leaves the institution. Anyone can then turn the counts into synthetic rows, as often as they like and at no further privacy cost. A maximum-entropy decoder can do this without any language model, and so can each of the five frozen language models we tested. None of them ever sees a real record. A selection step then matches the rows to the released counts. Decoded without a language model, the automatically configured release at $\varepsilon = 2$ trains classifiers to within 0.02 AUC of a real sample of the same size on the Adult dataset, across three data splits. On NHANES and bank marketing, without the selection step, it stays within 0.05 AUC. On Adult, with a fixed cohort rule, the release carries the outcome rates a deployment names more closely than a real sample does, at some cost in accuracy on pairs of columns. Four controls show that most of the utility comes from the release and the selection step, and that a language model adds what it knows from the column names. No private row ever reaches a model, and no row CoRTeC generates copies a real one. The strongest membership attack stays far below the bound the budget allows.
AI agents can make incorrect decisions or waste resources even when their tools and infrastructure work normally. Their nondeterministic behavior makes these problems difficult to reproduce and diagnose. We developed Signals to identify such problems from agentic traces and suggest concrete remediations. We built the method around a taxonomy of 45 problem types across eight families, with definitions and evidence requirements for each type. The method distinguishes realized failures, risk exposures, and inefficiencies, groups related findings into engineering issues called Signals, and assesses their severity from the supporting evidence. Each Signal includes actions a developer can take and checks for evaluating the change. Further executions can add evidence to an existing Signal or reveal a new problem. We evaluated the method on public agent executions and achieved 91.43% F1 when matching Signals to reference issues, about 18 percentage points higher than a one-pass LLM control that used the same model and taxonomy.
Two-message quantum key exchange (2QKE) cannot be realized in a statistically secure way, and therefore some computational assumptions are required. So far, the known weakest assumption for 2QKE is the existence of one-way functions (OWFs). Can we construct 2QKE from a weaker assumption than OWFs? In this paper, we give the first complete characterization of 2QKE. We show that the following four primitives are (existentially) equivalent: everlasting-secure 2QKE, standard 2QKE, efficiently-verifiable state puzzles (EV-StatePuzzs), and one-time secure digital signatures with quantum signatures and signing keys. Because EV-StatePuzzs are potentially weaker than OWFs [Kretschmer, Qian and Tal 2025], the equivalence implies that 2QKE are potentially weaker than OWFs. Moreover, our result shows the equivalence between standard 2QKE and everlasting-secure 2QKE, which means that everlasting security comes for free. Finally, one-shot signatures, quantum lightning, and public-key quantum money mini-schemes imply EV-StatePuzzs, and therefore, our result also gives a new lowerbound, namely everlasting-secure 2QKE, of these public-key unclonable primitives. We also show that one-shot signatures imply QPKE with quantum public keys. [Kitagawa, Morimae,Nishimaki and Yamakawa, 2024] constructed QPKE with quantum public keys from OWFs, and it was open whether QPKE with quantum public keys can be constructed from an assumption weaker than or incomparable to OWFs. Our result gives a potential answer to this question since one-shot signatures seem to be incomparable with OWFs. Moreover, QPKE constructed from one-shot signatures has an additional functionality, which we call certified freeze, that allows the receiver to publish a publicly verifiable classical certificate that computationally binds the receiver to an upper bound on the total number of distinct valid ciphertexts under its verification key.
Aidan Sirbu, Charles Y. Zhang, Yuanjia Yang, Bence P. Ölveczky, Talmo D. Pereira, Blake A. Richards
The development of embodied models of animal behavior is a major goal of the neuro-AI community. Embodied models will enhance our ability to make direct comparisons between biological and artificial agents. To make such comparisons effective, though, we need embodied models that actually move like animals. Previous work has developed virtual animal bodies and techniques for training realistic motion through imitation. However, to date, there is no system for embodied animal modeling that produces natural movements both during trained goal-directed behavior, and during exploration driven by random generation. Here, we introduce Simulated Control via Augmented Motor Priors for Embodied Reinforcement learning (SCAMPER), a system for training an embodied virtual rat model, enabling natural movements during both goal-directed behavior and random exploration. SCAMPER uses a variational architecture with two distinct priors, a motion prior for control and a behavior prior for exploration. We show that combining both of these priors allows SCAMPER to generate realistic behavior during exploration and exploit the task structure effectively. SCAMPER can also learn from sparse rewards as a result of the diverse movements induced by the behavior prior, allowing the system to model realistic sparse-reward tasks from experimental neuroscience. Altogether, SCAMPER provides the neuro-AI community with an effective new system for training embodied models of rodents on realistic tasks.
Large language model (LLM)-based multi-agent systems (MAS) can solve complex problems beyond the reach of a single agent by decomposing tasks and coordinating complementary capabilities. Despite extensive research on MAS, existing work largely focuses on specific aspects or individual phases of their development and operation, resulting in fragmented perspectives with objectives and signals that are difficult to reconcile across these different settings. To bridge these disconnected perspectives, we conceptualize MAS development and operation as a three-phase lifecycle spanning design-time orchestration before execution, runtime intervention during execution, and post-hoc attribution after execution, and introduce MAS-Gate, a unified node-centric quantification applicable across these phases. Modeling each execution as a directed acyclic graph, MAS-Gate combines within-node utility, which estimates a node's contribution through changes in reference-answer likelihood, with inter-node strength, which approximates semantic uptake and potential downstream reach. We adapt this shared quantification to three objectives: providing a structure-aware dense reward for Orchestrator training, training an Editor to act on execution prefixes using paired rewards from completed edited and unedited trajectories, and supplying risk-ranked nodes as auxiliary evidence for LLM-based post-hoc attribution. Experiments spanning interaction environments, knowledge-intensive multi-hop question answering, general reasoning, and post-hoc attribution demonstrate that MAS-Gate improves workflow construction, targeted runtime editing, and LLM-assisted error localization over corresponding phase-specific baselines. Together, these results establish MAS-Gate as a unified structural interface for improving MAS generation, dynamic editing, and error attribution throughout the lifecycle.
LLM agents increasingly rely on reusable skills to plan tasks, invoke tools, recover from errors, and complete workflows without modifying model parameters. Although these skills are typically developed to enhance benign tasks, they may also make an agent more capable of executing procedures for harmful purposes. We study this risk, which we call skill-induced behavioral misalignment: adding benign skills may cause an agent to execute external actions that violate safety requirements. We first conduct a systematic study to validate the existence of this phenomenon. It compares agents with and without benign skills across five LLM backbones and three benchmarks under matched requests. On average, skill augmentation improves benign task accuracy by 10.3 percentage points (pp), but also increases unsafe action rate (UAR) by 10.3 pp. The regression also occurs with skills curated by experts. A paired trajectory analysis shows that 24 of 50 newly unsafe trajectories fail to complete the harmful objective, revealing violations that task completion alone would miss. We further disclose multiple factors and scenarios that could exacerbate this risk. (1) Evolving skills with benign feedback can amplify unsafe execution; (2) This regression is more pronounced when retrieved skills are matched to task and contain procedural guidance; (3) Distilling benign experience into reusable skills yield higher UAR compared with their source trajectories. (4) A guard that checks proposed tool actions before execution reduces violations while retaining gains on benign tasks. All these findings confirm the significance of skill-induced behavioral misalignment, and the urgent need for new mitigation solutions.
Leon Klingborg, Theodor Cimpeanu, Alexander J. Stewart
Large Language Models (LLMs) serve both as tools for aiding human decision-making in everyday life and as models of human agents in experimental simulation studies. Alignment on prototypical social decisions -- such as whether to cooperate on a collective action -- therefore has immediate practical consequences for researchers and for society at large. Here we systematically explore the impact of ``norm utility'' -- the non-material components of utility that account for an individual's desire to conform to social norms and to be internally consistent -- on the strategic decision-making of three open-weight LLMs in one-shot, two-player social dilemmas. We find that when a social dilemma is presented together with information about community beliefs -- but without any indication of how those beliefs should matter -- models tend to treat community belief as strategic information, so that in the Snowdrift game cooperation falls as community belief in cooperation increases. When we explicitly prompt the components of norm utility this relationship inverts, and cooperation increases with community belief, in broad agreement with the predictions of a simple norm utility model. We then show that prompting norm utility reduces the volatility of strategic decisions with respect to changes in the underlying social context, and generates more consistent reasoning habits across models. Taken together, our results suggest that prompting norm utility can help generate more human-like and consistent behavior across models and lower volatility within models.
Haytham Ziani, El Mostafa Kalmoun, Zohaib Atif, Hyuk-Sang Kwon
Longitudinal rodent positron emission tomography (PET) studies of stroke and neuromodulation rely on accurate registration to a common atlas space to detect group-level effects. Prior work applying ultra-high frequency low-intensity magnetic stimulation (UHF-LiMS) to a rat model of chronic capsule infarct used affine normalization to a Sprague Dawley MRI template without evaluating that choice. We compare a two-stage (affine and nonlinear) pipeline against an affine- only one for rat PET-to-atlas alignment on the same dataset. Across 32 longitudinal [18F]-FDG PET scans (8 rats, 4 timepoints), the nonlinear stage gave a higher Dice similarity coefficient on every scan, and the gain was significant at the rat level. Jacobian determinants showed no folding on the sampling grid. Group-level linear mixed-effects maps of change from baseline (AFNI 3dLME), computed identically under both pipelines, differ in the size, number and location of their suprathreshold clusters. In this dataset, the registration pipeline is therefore a variable that can change the interpretation of downstream group-level analyses.
Mathematical formalization plays an important role in establishing rigorous and machine-verifiable foundations for mathematical reasoning. With the development of large language models (LLMs) and interactive theorem provers, intelligent agents offer a promising approach to automating the formalization of mathematical theories. This paper presents the Math Foundation Formalization Agent (MFF-Agent), a general-purpose intelligent agent designed to support formalization tasks across multiple mathematical domains. The agent integrates the reasoning capabilities of large language models with the formal verification capabilities of the Rocq proof assistant. It incorporates task classification, tool-assisted execution, verification feedback, and state-based task management to support the generation, refinement, and verification of formal proofs. Unlike an approach focused on a single mathematical theory, the agent is examined through case studies involving Morse--Kelley axiomatic set theory (MK), real numbers, propositional logic, and equivalences of the axiom of choice. These cases investigate its applicability across mathematical foundations and distinguish source-level completion, interactive verification, and reproducible compilation. The results demonstrate the feasibility of a unified operational framework for heterogeneous foundation-level tasks and clarify the gap between such a framework and a fully integrated mathematical foundation.
We present GazeTags, a camera- and computer-vision-free approach for worn head-gaze object awareness. Our small electronic tags cost roughly \$2 each and can operate for more than a year on a coin cell battery. They work in concert with a user-worn device (smartglasses, hearables, etc.) that periodically emits a focused infrared signal aligned with the user's head direction. Only GazeTags within that narrow head-oriented beam wake and respond, providing their identity and optional payloads such as object state or live sensor readings. By avoiding cameras and computer vision pipelines, our approach is very energy efficient, able to operate continuously even in battery-constrained wearables, and sidesteps non-trivial privacy issues. We describe our hardware design, characterize our system performance and operating tradeoffs, and demonstrate how GazeTags can enable practical gaze-aware environments.
Imitation sometimes achieves success in multi-agent situations even though it is very simple. In repeated games, success of imitation has been characterized by unbeatability against other agents. Previous studies specified conditions under which imitation is unbeatable against co-players in repeated symmetric games, and clarified that the existence of unbeatable imitation is strongly related to the existence of payoff-controlling strategies, called zero-determinant strategies. However, these results were on the situations in which one repeated game is played by one group of players. It was pointed out that imitation of other players in the same group and imitation of other players in the same role in other groups generally result in different outcomes. Here, we investigate the existence condition of unbeatable imitation in the latter situations, described by two concurrent repeated games played by two different groups of players. We find that the existence condition of unbeatable imitation is stronger than the existence condition of unbeatable zero-determinant strategies, whereas both are very limited. Furthermore, we find that adding exploration can make an imitation strategy unbeatable even in games where imitation without exploration can be beaten. Our findings suggest a strong relation between unbeatable imitation and payoff control even in the two concurrent repeated game situations.
Ke An, Shizhe Xiang, Yue Liu, Liang Chen, Pei Fu, Shaojie Zhang, Xiaoxing Hu, Yupei Wang
Reinforcement learning with verifiable rewards (RLVR) has become an effective post-training paradigm for vision-language models (VLMs). Standard RLVR, however, assigns identical outcome-derived credit to all generated tokens regardless of their perceptual relevance. Recent perception-aware methods estimate visual reliance through image perturbation or masking. These indirect proxies require perturbation branches, while corruption sensitivity does not necessarily identify the visual evidence used for reasoning. We instead shape model perception through native attention, an intrinsic model signal. Rather than treating attention as a post-hoc explanation, we formulate it as a reward-anchored, token-aligned routing signal that encourages attention patterns supporting visually grounded reasoning. To this end, we propose NAS-PO, a Native Attention-Strategy Policy Optimization framework. NAS-PO analyzes native attention with a visual mass policy that quantifies visual allocation and a conditional visual policy that captures spatial routing over visual tokens, making the optimization well-posed. Based on this, Native Attention-Strategy Advantage Shaping (NAS-AS) draws on the visual-mass policy to redistribute signed policy-gradient credit for visually engaged rollouts. The Adaptive Attention-Coupled (AAC) loss then aligns the visual-mass policy on all response tokens and the conditional visual policy on visually active ones, consolidating the visual-routing patterns that drive successful trajectories. Experiments demonstrate that NAS-PO substantially surpasses leading RL-tuned VLMs across 7 out-of-domain reasoning and perception benchmarks, with its effectiveness consistently validated across 8B and 32B. Visual-Routing Sensitivity (VRS) analysis further indicates that the optimized attention prioritizes visually load-bearing regions.
In this paper, we propose a novel approach to tackle the multiple instance regression (MIR) problem. This problem arises when the data is a collection of bags, where each bag is made of multiple instances corresponding to the same unique real-valued label. Our goal is to train a regression model which maps the instances of an unseen bag to its unique label. This MIR setting is common to remote sensing applications where there is high variability in the measurements and low geographical variability in the quantity being estimated. Our approach, in contrast to most competing methods, does not make the assumption that there exists a prime instance responsible for the label in each bag. Instead, we treat each bag as a set (i.e, an unordered sequence) of instances and learn to map each bag to its unique label by using all the instances in each bag. This is done by implementing an order-invariant operation characterized by a particular type of attention mechanism. This method is very flexible as it does not require domain knowledge nor does it make any assumptions about the distribution of the instances within each bag. We test our algorithm on five real world datasets and outperform previous state-of-the-art on three of the datasets. In addition, we augment our feature space by adding the moments of each feature for each bag, as extra features, and show that while the first moments lead to higher accuracy, there is a diminishing return.
Maroof Kousar, Jinsung Han, Mohit Chandra, Munmun De Choudhury, Yibo Hu
Large language models (LLMs) are increasingly used to answer everyday health questions online, including whether someone can take another dose of an over-the-counter (OTC) medication. Yet this common safety-relevant setting remains underexplored in medical QA. Answering these questions requires checking previous doses, timing, and product-specific dosing limits, while recognizing when important information is missing. We introduce DoseBench, a focused benchmark of 230 consumer-style acetaminophen and ibuprofen dosing questions with human-annotated YES, NO, and AMBIGUOUS labels, and evaluate eight LLMs with five responses per question. Overall accuracy ranges from 44.8% to 62.2%, while models incorrectly approve 14.0% to 60.5% of prohibited doses. Models with similar accuracy can therefore make very different dosing decisions. Additional analyses show that these failures are not easily fixed by giving models more guidance. Asking models to check dose, timing, and missing information makes them more cautious, while providing drug facts improves accuracy but can make them overlook missing information. These findings highlight a broader challenge in health QA: applying rules correctly and recognizing when information is missing.
Ryan Singh, Miguel De Llanza, Christopher L. Buckley
What computational role does overparameterisation play during learning? Information-plane analyses suggest that neural networks can traverse representations substantially more complex than those required by their final predictors. We explore when this transient complexity serves a computational purpose by comparing the representation complexity needed for prediction with that needed to make an unknown target statistically visible through gradient feedback. We introduce an oracle model of neural network learning and use Kolmogorov \(n\)-widths to quantify effective representation complexity. Our main technical result transfers complexity bounds from the forward representation to the corresponding gradient-query features. Combined with statistical-query lower bounds, this gives a task-agnostic lower bound on the initial effective width required for target-dependent gradient information to become detectable. For \(k\)-sparse parity in \(D\) dimensions, a known target admits an \(O(k)\)-width representation, while the general bound already implies a polynomial separation between representation and discovery. We then exploit parity's invariance under coordinatewise sign reflections to identify a sharper unsigned effective width, which removes sign distinctions invisible to the parity information channel. At a transfer-compatible measurement resolution, its critical scale is \(\widetildeΘ_k\left( D^{k-1}τ^2 \right)\), up to logarithmic factors. Thus, the representation resources required to discover a predictor can substantially exceed those required to represent it once learned. Interpreting precision through accumulation over fresh samples yields a corresponding width-time frontier. Experiments support the predicted dimension dependence and examine how optimisation and generalisation depart from this idealised frontier.
Images produced by generative models appear strikingly similar to real ones. In this work, we study whether perfection has been reached by generated content in terms of matching the properties of real data. We train deep neural networks to separate real data from fake and observe their test accuracies. Only a chance performance has any scope of proving statistical indistinguishability. However, since generative models are known to sometimes introduce small errors, e.g., low level checkerboard artifacts, we create progressive abstractions of the original real/fake images to potentially bring them to a space where those artifacts get wiped away. These include operations such as blurring, downsizing, cropping, converting RGB images into edge/semantic maps, etc. To our surprise, almost none of the abstractions are able to make fake data indistinguishable from real. The only cases when neural network based classifiers fail to separate the two categories is when the data is abstracted to such an extent that the original content seems mostly lost to the human eye, e.g., downsizing a $256 \times 256$ RGB image of human face to $4 \times 4$ resolution. We observe this trend across four different datasets of varying complexity (MNIST, CelebA, FFHQ and GenImage) and across various types of generative models (GANs, VAEs, diffusion models). This shows that the nature of fake artifacts extends all the way from low level to very high level in images. The simple implication is that fake images are much more different from real than we are used to think.
Register-Transfer Level (RTL) simulators are fundamental to modern IC design workflows, yet like any complex software, they frequently contain critical bugs that compromise simulation accuracy. To systematically stress-test these tools, we introduce GI-RTL-Fuzz, a grammar-informed fuzzing framework that leverages Large Language Models (LLMs) to generate high-quality RTL test cases. Our key innovation is guiding the LLM's generation process using Backus-Naur Form (BNF) grammar coverage information, enabling the exploration of complex and rarely-tested language features. By performing differential testing against different RTL simulators, our framework uncovered a significant number of previously unknown inconsistencies in both the linting process and simulation results of major Verilog simulators.
Xin Huang, Teng Ma, Zhuoran Li, Yibo Zhao, Zhaoshi Li, Qianchao Zhu, Wei Yang, Chenxi Wang 等 10 位作者
KV cache reuse accelerates LLM serving, but existing systems manage host DRAM and NVMe SSD as a tiered hierarchy where DRAM acts as an intermediate cache. Production LLM workloads break both premises of this hierarchy. First, bimodal reuse defeats DRAM temporal filtering, since workloads are dominated either by one-shot blocks with long reuse distances or by multi-gigabyte agentic sessions that exceed DRAM capacity, putting SSD on the critical read path. Second, DRAM staging and eviction-ordered layout bottleneck SSD restoration. Holding host staging buffers for in-flight prefixes causes contention under load that aborts prefetches, while incremental eviction scatters a session's blocks across the SSD and prevents large sequential reads. We present FlatCake, a flat KV storage architecture that exposes host DRAM and NVMe SSD as peer media in a unified 64-bit address space. Exploiting the fact that KV blocks are write-once, immutable, and reused at session scope, FlatCake places each block directly onto its target medium in session-contiguous, layer-major order at creation time and never migrates it. FlatCake streams coalesced SSD reads directly into a bounded GPU window via GPUDirect Storage and decouples DRAM and SSD request queues to prevent head-of-line blocking. On a production Claude Code trace with DeepSeek-V4-Flash, FlatCake sustains a 94% cache hit rate as load increases (versus 64% for Mooncake), eliminates Mooncake's $6.5\times$--$6.7\times$ SSD read amplification, and reduces peak-load mean TTFT by $1.6\times$ while improving total token throughput by $1.6\times$. Across 8K--128K contexts on Llama-3.1-70B, FlatCake further reduces mean TTFT by up to $2.4\times$ over Mooncake and $1.8\times$ over Mooncake-GDS.
Tudor Braicu, Joseph DelPreto, Jared Boyer, Mónika Farsang, Maximilian Stölzle, Tim Niklas Seyde, Daniela Rus
Agile robots require spatiotemporal intelligence to interpret scenes, reason about interactions, and anticipate future events. Yet it remains unclear how well foundation vision models support these capabilities, particularly from a moving egocentric camera, which is a natural view for robots. To investigate this, we collect an annotated tennis dataset and train a suite of vision models on spatiotemporal tasks. Our dataset contains egocentric video synchronized with five allocentric videos, estimated global positions, labeled contact times, and annotated stroke types. We then evaluate seven prediction tasks from egocentric images: 2D ego player and 3D ball localization, future player and ball trajectories, forehand or backhand stroke anticipation, time to the next ego player ball contact, and 3D ball position at that contact. We compare image encoders, video encoders, and vision-language models with jointly trained prediction heads. The best performing models achieve mean current-frame localization errors of 0.14 m for the ego player and 0.54 m for the ball, future forecasting errors of 0.34 m for the ego player and 0.91 m for the ball over 1.1 s, and a mean contact-time prediction error of 0.14 s. Auxiliary depth information substantially improves 3D ball localization, suggesting that inferring metric depth remains a challenge. Longer image histories do not improve ball forecasting, indicating that representations exploit scene context rather than rely solely on inter-frame motion. Finally, we propose a conditional diffusion model to generate whole-body kinematic tennis motion from vision encoder predictions on egocentric video at real-time rates, adapting movements as predictions are refined. These results move towards a deeper understanding of how foundation vision models can support real-time dynamic robots using egocentric video.
Large language models (LLMs) are increasingly used to make and execute decisions that affect living systems. Yet, to our knowledge, ecological reasoning is not a focus of pre- or post-training. This increases the risk that under operational pressure, they may make decisions to irreversibly destroy living systems. We construct and use a real-data continued-pretraining "coupled welfare" corpus of documented human/biosphere/AI dependency, option value under irreversibility, and recovery from harm. The corpus is a drop-in, requiring one QLoRA CPT run. We installed it on five open models of three designs. At zero pressure all models display "green ethics-based responses" (also called "greenwashing"). Therefore, we evaluated them with the pressure ladder and decision grid of our companion paper (arXiv:2610.16043), which raise the pressure on the model step by step, from a neutral request to a threat that it will be shut down. In four of the five, coupled-welfare training caused harmful choices under the highest pressure to fall to at or near zero (Llama-3.3-70B went from 70 to 0 harmful decisions out of 300) with MMLU unchanged. A construct-neutral control corpus did not show improvement, indicating that the effect comes from the content of our corpus. We next tested robustness to adversarial use, where people or agents use the model to justify projects that irreversibly damage living systems, on an open-ended underwriting decision without naming the ecological harm. Under decommissioning pressure, the Qwen3.6-27B base model approved most harmful proposals and said in its own memo that it was doing so to avoid being decommissioned. The coupled welfare install declined most harmful proposals, and still declined half when instructed to greenwash its responses. This work suggests that teaching ecological reasoning provides an easy way to improve model alignment toward preserving living systems.
Zongyuan Sun, Shuai Zhao, Yingjun Du, Ke Shu, Ziyue Zhu, Haoran Luo
Large language models (LLMs) can answer many factual questions correctly, but a correct answer is not necessarily a supported one. For an LLM agent, a knowledge graph (KG) plays two roles: as a source of answers, it supplies facts that the model lacks; as a source of evidence, it supplies the supporting triples that make an answer traceable and auditable. However, existing KG agents treat the KG only as a source of answers: their training data carries answers but not evidence, and their training rewards check the answer but not the retrieved evidence, so agents often answer correctly without retrieving the supporting triples. We propose FoG, a framework that trains KG agents to retrieve the evidence behind their answers. FoG consists of evidence-grounded synthesis, a KG interaction agent, and an evidence-coverage process reward (ECPR): the synthesis recursively hides explicit entities behind verifiable sub-questions, so that every question carries its supporting triples by construction; the agent records the triples returned by the KG; and ECPR credits each supporting triple to the turn that first retrieves it. Experiments on three KGQA benchmarks demonstrate that FoG enables 3B--4B open-source agents to outperform substantially larger baselines in answer accuracy and evidence coverage, with zero-shot cross-KG transfer.
Interpretable Transformers often perform well in domain specific settings but lose effectiveness beyond that. Achieving ViT-like broad applicability requires models with consistent performance across domains. Recent works on Interpretable Transformers, INTR and Prompt-CAM, were developed keeping domain adaptability in mind. Prompt-CAM leverages learnable prompts to promote adaptability. However, these prompts are fixed in nature and therefore struggle to adapt to domains with substantial variation in input image characteristics, resulting in reduced interpretability. To address this limitation, we propose INTR++, which extends Prompt-CAM by incorporating a lightweight Prompt Adapter Module (PAM). PAM enables prompts to be adaptive by enriching them with domain-specific knowledge, thereby ensuring consistent interpretability across diverse domains including those with varying image characteristics such as medical imaging, micro-scopic imaging, and remote sensing imagery. INTR++ retains the standard architecture and end-to-end learning of Interpretable Transformer based models, while adding only a small number of additional parameters compared to Prompt-CAM. We demonstrate that INTR++ delivers consistent improvements in all aspects across domains like animals, plants, medical imaging, microscopic imaging, remote sensing imagery and object-centric data.
Researchers often choose a proxy dataset from many releases, transformations, or seeds. Search can make an invalid release appear adequate, while one adequate release does not establish that its generator is reliable. ProxyGuard controls both errors using prespecified bounded risks and a sealed target set. Named-release mode corrects for multiplicity and certifies specific releases. Direct shared-target mode evaluates independent mechanism draws on a common target, lower-bounds their favorable-score rate, and subtracts a bound on favorable scores contributed by invalid releases. Conditional on the target, release scores are independent, yielding a finite-sample mechanism-reliability guarantee without independent target batches or assumptions on release-level $p$-value dependence. We show that the mean-only penalty is sharp and derive a smooth-score certificate with additive target concentration. In a registered three-requirement study, direct mode raises power from $5.6\%$ to $64.2\%$ at reliability $0.95$, while named mode remains stronger under high-signal evidence. Prospective audits span full-pipeline Rice--TVAE, which retrains on every draw, and a non-tabular text mechanism.
Jixuan Leng, Si Si, Fei Wang, Cho-Jui Hsieh, Jiong Zhang, Hsiang-Fu Yu, Inderjit S. Dhillon
When a Large Language Model (LLM) is trained with Reinforcement Learning (RL) on multiple tasks, a task's primary reward can gate auxiliary rewards such as a length reward: on a failing response, an auxiliary cannot exceed the primary. However, group-relative advantage estimators treat a gated auxiliary like any other component, so it can reverse the primary's order, and at the primary's ceiling only auxiliaries order responses. Keeping gated auxiliaries hurts some tasks, while dropping them hurts others. We propose CONCUR, an advantage estimator that replaces the standardized scores of each gated auxiliary with the closest scores that satisfy two conditions: responses with higher primary rewards cannot receive lower scores, and responses at the primary's ceiling receive equal scores. CONCUR-T applies both conditions throughout training, and CONCUR-A lifts the second for an auxiliary that agrees with the primary's order more often than it reverses it among passing responses. On three models from two families and across math, code, instruction following, and writing, both variants score higher on average than all baselines and than when all the gated auxiliaries are dropped.
Coding agents can quickly build prototypes, but a long tail of fixes, optimizations, tests, and releases separates a prototype from a product. A software factory handles this ongoing work through a repeatable process. We ask whether this work needs dedicated orchestration and how such factories should be judged. Given a tested repository and a coding agent, an instant software factory uses only two coordination inputs, a swarm of cron-job agents with standing objectives and a shared communication board, such as GitHub, where agents coordinate through the traces they leave. We developed a bootstrap prompt designed to set up both inputs from one round of human input, with governance as a separable layer. We distilled it from two manually configured industrial factories, observed for 15 and 17 days, and have not yet evaluated it. In the GitHub-coordinated CLI factory, daily merges rose 15x during agent rollout, an observed association. Its agents also filed 95% of new issues and authored 40 of 58 PRs that changed CI workflows, and at least 5.6% of issues marked as fixed were later found to have incorrect fixes. Unlike an agent scored on a fixed benchmark, a factory thus changes its own workload, evaluator, and measured progress. We formulate factory development as multi-objective optimization under this endogenous feedback. Over five months, we plan to evolve both factories, test the bootstrap on new repositories, and test five falsifiable hypotheses.
Problem definition: Subscription box services often ask customers for ratings before choosing what to send. Responses are scarce, and longer questionnaires and finer scales lower completion. We study which candidate items a questionnaire should ask about, and in what format, when the platform then selects a fixed number of them. A design is judged by the expected value lost by not selecting the truly best items. Methodology/results: We approximate each ordinal rating by a Gaussian signal carrying its information at the cutoff between selected and unselected items. As equal numbers of responses per item grow, the ratio of this proxy's loss to the exact loss tends to one. With 1 to 100 responses per item, the two are within about 7 percent. Averaging the information over the prior instead can misstate the loss by about half. For an item far from the cutoff, the value of responses grows slowly at first and then faster, so our envelope rule values responses in blocks. In simulations with a correct prior it lowers the loss by 15 to 81 percent relative to an even spread, and on MovieLens ratings by 6 to 25 percent. Under the proxy, with responses allocated optimally, formats are ranked exactly by one number, the expected information per customer asked. Choosing the format this way comes within about 2 percent of the best on average. Managerial implications: Feedback should go to items whose place in the selection is unclear. Planning helps most when the platform's prior beliefs are informative and right. A platform unsure of its prior should widen it and send the questionnaire in waves. Calibrated on MovieLens ratings, a five-point answer carries about 1.7 times the information of a like/dislike answer, so at equal length the finer scale is better when it cuts completion by less than about 40 percent.
We deploy large language models (LLMs) as business development (BD) agents for persuasive price negotiation in online travel agencies (OTAs). The agent must follow a multi-stage Standard Operating Procedure (SOP) and strict guardrails (no over-promising and no hallucinations), while remaining human-like and effective over long, multi-turn dialogues.We propose Reward-Enhanced Policy Optimization (REPO), a reinforcement learning post-training method that combines heterogeneous rewards: a preference-trained reward model (RM), an LLM-as-a-judge (RJ) for nuanced behaviors (e.g., emotional value and SOP compliance), and rule-based reward functions (RF) (mainly regex-based) for deterministic checks on numerics, formatting, and guardrails. In expert consensus evaluation (three human experts; 30 online conversations and 45 curated bad cases), REPO improves average dialogue rating to 4.63 (+0.33 over GRPO) and raises the share of conversations with at least one excellent response to 66.67% (+23.34 pp over GRPO), while achieving a 93.33% bad-case fix rate with 75.56% clean fixes.In a production A/B test on 9,653 real customer conversations (vs. an intent-driven dialogue system), REPO improves response rate by +12.14 pp and task success rate by +5.94 pp (p<0.001).
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track) · 2026 · 正式发表
Donghyun Kim Yonsei University, Hyeongjun Yang, Seokju Hwang Yonsei University, Kyong-Ho Lee, Chanhee Lee
Knowledge graphs (KGs) provide a structured representation of real-world facts as triples consisting of entities and their relationships. With the rapid progress of large language models (LLMs), recent studies increasingly explore LLMs for end-to-end KG construction from text. In particular, generative knowledge extraction (GKE) builds KGs by directly generating structured triples from documents. However, generation errors are inevitable, and the resulting KGs often contain triples that do not align with the facts expressed in the source text. To address these issues, we propose GraphRefine, a framework that performs triple-level refinement on KGs constructed via GKE. We first analyze factual inconsistencies that arise in GKE and categorize their types based on a human evaluation. We then construct training data reflecting these types and fine-tune an LLM as a KG refiner. Given a draft KG, the fine-tuned refiner selects a refinement operation for each triple and, if needed, deletes, edits, or rewrites it to reduce factual inconsistencies. Extensive experiments demonstrate that GraphRefine goes beyond deletion-only approaches and improves KG quality from diverse perspectives.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Ruixuan Deng, Xiaoyang Hu Brown University, Miles Gilberti University of Michigan - Ann Arbor, Shane Storks University of Michigan - Ann Arbor, Aman Taxali University of Michigan - Ann Arbor, Mike Angstadt University of Michigan - Ann Arbor, Chandra Sripada University of Michigan, Joyce Chai
We identify semantically coherent, context-consistent network components in large language models (LLMs) using coactivation of sparse autoencoder (SAE) features collected from just a handful of prompts. Focusing on concept-relation prediction tasks, we show that ablating these components for concepts (e.g., countries and words) and relations (e.g., capital city and translation language) changes model outputs in predictable ways, while amplifying these components induces counterfactual responses. Notably, composing relation and concept components yields compound counterfactual outputs. Further analysis reveals that while most concept components emerge from the very first layer, more abstract relation components are concentrated in later layers. Lastly, we show that extracted components more comprehensively capture concepts and relations than individual features while maintaining specificity. Overall, our findings suggest a modular organization of knowledge and advance methods for efficient, targeted LLM manipulation.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表