Jev is a commercial "System One" model from TypeSafe AI that does not generate text: given a state and typed questions, it returns a choice from fixed options, a position on a rubric, or the probability that a statement is true, with probabilities described as calibrated. Such models target small decisions in information access pipelines. We evaluate Jev (jev-1.13.0) zero-shot on 37 datasets, from classification and routing to moderation and rubric scoring, with one frozen template per dataset and full evaluation splits: 346,009 requests for under US$10. For reference, we score Qwen3.8-27B and Gemma-4-E4B on identical requests via their exact option probabilities. Jev reaches 95-99% accuracy on IMDB, SST-2, HellaSwag and ARC, beats Qwen on 27 of 37 datasets, with none of Qwen's leads outside the bootstrap intervals, and Gemma on all 37, and its choice probabilities are well calibrated and support selective prediction. All three models degrade on low-resource languages, fine-grained or noisy labels, and rubric-based judgments, and binary probabilities need tuned thresholds, which raise micro-F1 on UNFAIR-ToS from 0.50 to 0.75. Jev answers MMLU's calculation-heavy questions more accurately than other MMLU questions, unlike both open models; probes rule out shallow memorization but not memorized question-answer pairs. Extending the protocol to sequential decisions, where the model chooses the next action at every step, Jev completes 35% of ALFWorld's unseen tasks, 42% of 90 TextWorld games and 21% of BabyAI-Text missions, chooses the optimal Blackjack action in 90% of decisions, and on FrozenLake almost never moves away from the goal to avoid holes. It often repeats actions that had no effect. Qwen completes more ALFWorld and TextWorld games than Jev (54.5% and 51.1%) but shares its failures on BabyAI-Text and FrozenLake. We release the code, harness and all raw responses.
Subhashini Venugopalan, Jenny Tan, Taylor Roper, Jimmy Tobin, Anton Kast, Alicia Martin, Sam Sepah, Amy Pavel
Audio descriptions (AD) are essential for making visual media accessible to blind and low-vision (BLV) users. While Multimodal Large Language Models (MLLMs) offer a scalable solution for generating audio descriptions on-demand, their performance on ``inline'' audio descriptions -- which must fit within existing silences in a video -- remains under-explored, particularly for diverse user-generated content (UGC). We present a comprehensive investigation into MLLM-generated inline audio descriptions. First, we introduce MTG-AD, a professionally annotated dataset of 36 hours of videos, each averaging $\sim$5 mins. and 8k+ descriptions in total, co-created by paired blind and sighted AD professionals to provide a high-quality benchmark for this task. Second, we propose MTG-AD-Eval, an evaluation framework based on audio description expert guidelines that provides fine-grained, reference-free actionable metrics for model improvement. Crucially, our framework evaluates the end-to-end task: it renders the generated scripts into audio tracks to assess their fit within the original video's natural silences, minimizing disruptive overlaps. Our analysis across frontier MLLMs reveals that while models accurately describe visual events and substantially reduce speech overlaps, gaps persist in narrative pacing and flow. Finally, we empirically validate that our automated and reference-free metrics correlate significantly with human quality ratings from sighted raters and 7 BLV participants, finding that while users narrowly prefer human-authored audio descriptions, they find MLLM-generated audio descriptions highly beneficial for comprehension. We provide actionable system design guidance and release MTG-AD at https://hf.co/datasets/google/mind-the-gap-inline-ad and MTG-AD-Eval at https://github.com/google/mtg-ad.
Privacy accounting bounds the cumulative privacy cost of Differentially Private Stochastic Gradient Descent but does not determine how to allocate a fixed budget across iterations. We formulate the complete Gaussian noise-variance schedule as an open-loop control under a composed zero-Concentrated Differential Privacy constraint, explicitly accounting for schedule selection. For expected quadratic terminal tracking, we prove that every local minimizer exhausts the certified budget, yielding an exact-budget simplex formulation. We introduce Privacy-Budget Adjoint Mirror Descent (PB-AMD), combining pathwise adjoint gradients with entropic updates that preserve positivity, charge floors, and the prescribed budget. Under regularity and coercivity assumptions, we establish existence and budget stability of optimal schedules, design-to-training utility stability, and deterministic, stochastic, and sample-average stationarity guarantees. A linear-quadratic reduction yields an explicit square-root allocation law. Experiments demonstrate tracking improvements over several allocation rules and predictive improvements over uniform allocation in a learning-rate restart setting, where learning-rate allocation performs similarly. Comparisons at a common stationarity tolerance show lower schedule-optimization times than projected gradient descent on the tested problems. The framework connects certified privacy accounting with optimization through nonlinear training dynamics.
Chenxiao Fan, Chongming Gao, Gangyi Zhang, Yaxin Gong, Leyang Shen, Jiakai Wang, Dong Wang, Yang Liu 等 12 位作者
Complex information seeking on the Web often requires collecting evidence from multiple documents and reasoning across them, as in multi-hop question answering. Multi-agent systems address such tasks by dividing work among specialized agents, where upstream agents produce intermediate answers and supporting evidence that downstream agents build upon. However, upstream answers vary in correctness, and most systems provide no quantitative signal about it, leaving the downstream agent to judge whether to trust or verify what it receives based solely on the textual evidence. In this paper, we propose MASCAL. Each agent learns to produce a correctness probability for its answer by jointly optimizing task performance and confidence calibration in reinforcement learning; the upstream agent passes this probability at handoff, giving the downstream agent a quantitative basis for its decisions, and the final answer also carries a calibrated probability, quantifying the reliability of the system output. To guide the downstream agent in using this signal during training, we introduce confidence-aware weighting, which increases the learning weight at positions responsive to the upstream probability, so that the agent learns more from decisions where it actively uses the signal. Across three collaboration architectures spanning Web retrieval and reasoning tasks, MASCAL improves task accuracy and calibration quality for both agents, and generalizes to deeper collaboration chains and different model scales. Intervention analysis confirms that upstream confidence affects downstream decision behavior, and the learned final confidence supports selective answering and targeted rechecking.
Yansong Bai, Yunzhi Zhuge, Xinzhuo Yu, Lu Zhang, Jin Zhan, Pingping Zhang, Huchuan Lu
Domain generalized semantic segmentation is challenging because it requires accurate dense prediction on unseen domains without any access to target-domain data during training. Existing methods improve transferability through data diversification, invariant learning, or direct use of foundation models, but they often struggle to simultaneously suppress domain-sensitive interference and preserve fine-grained spatial cues. In this work, we propose SIFR, a frequency-aware dual-stream framework that tackles domain shift at both the fusion and decoding stages. SIFR uses spectral modeling as a unified principle to regulate both heterogeneous prior interaction and dense decoder refinement. Instead of directly fusing frozen CLIP and DINOv2 features, SIFR performs spectral-guided invariant fusion to filter out domain-sensitive responses while preserving complementary semantic and structural information. In addition, SIFR incorporates a spectral decoder refinement module that stabilizes multi-scale decoder features through low-frequency statistics-based normalization and adaptive frequency selection, thereby reducing the propagation of residual domain bias in dense prediction. This design enables SIFR to jointly improve domain robustness and segmentation precision under severe distribution shifts. Extensive experiments on multiple DGSS benchmarks, including synthetic-to-real, real-to-real, and adverse-weather settings, demonstrate consistent improvements over previous state-of-the-art methods.
As language models (LMs) have improved, an increasingly common means of evaluating their linguistic competence is through explicit prompting (e.g., "Which sentence is more acceptable?"). Prior work indicates that such prompting underperforms compared to standard token-probability-based methods in smaller models, but improves in larger models. We test this in a new domain: semantic focus-sensitivity, responsible for contrasts like "Lucy only SWAM yesterday" vs. "Lucy only swam YESTERDAY". Testing models of various sizes using text-based focus markers like boldface and italics, we find universal above-chance performance for token probabilities. Prompting results mostly align with previous findings, but in some larger models they outperform probabilities by a wide margin. We propose that this is because LMs are not prone to produce out-of-the-blue focus marking, which muddies results from token probabilities in a way that does not affect prompting. We test this by replacing standard focus markers with randomly generated ones; token probabilities still beat chance but show significantly greater performance decrease compared to prompting. These results suggest that metalinguistic prompting, while non-trivial for LMs, can also be more robust when testing stylistically atypical data. We argue that this has important consequences for the assessment of LMs' language competence, even if LMs do not perform metalinguistic judgments through human-like introspection.
Elias Sandmann, Sebastian Lapuschkin, Wojciech Samek
Recent mechanistic work has uncovered learned algorithms within neural networks, from modular arithmetic to search and planning in game-playing agents. But does algorithmic structure guarantee algorithmic behavior? We investigate this in Leela Chess Zero, the strongest neural chess engine, where prior work identified learned look-ahead. By extending the logit lens to its move-selecting policy network, we discover that correct puzzle solutions--including immediate checkmates--often appear in intermediate layers but are systematically overridden in the final output, a phenomenon we term "forgotten puzzles". Replicating prior analyses on these positions, we find that look-ahead operates normally--future moves of the correct continuation are represented, causally important, and linearly decodable--ruling out a failure of the algorithm. Instead, late layers increasingly shift toward safe play over aggression. To test whether this shift drives the override, we steer the model against these preferences and recover 71.7% of forgotten puzzles, providing causal evidence that safety priors override algorithmically computed solutions. These findings show that algorithmic structure does not guarantee algorithmic behavior: a model can internally solve a problem and still output the wrong answer.
Clients observing different parts of a prediction task may learn locally equivalent models with incompatible behavior outside their observed support. We study whether redundant models permit local low-loss paths toward a compatible shared model. CovFedAvg retains local models and curve controls on clients, fits a common endpoint through aggregated gradient contributions, and selects an actual candidate model using distributed training losses. No centralized calibration examples or client-to-client sample exchange are required. Exact selection guarantees no higher evaluated training loss than the previous model and the current-round FedAvg candidate; this is not a global-optimality, stationarity, generalization, or differential-privacy guarantee. We state conditional client-compatibility and approximate-selection bounds and retain corrected transformation identities and a separate full-gradient safeguard theorem for reference algorithms. A complementary input-support example recovers the shared absolute-value regression, while matched direct fine-tuning also solves it. Across a synthetic support task, recorded activity-recognition subjects, and participating and future gas-sensor batches, we evaluate five methods, five seeds, two network widths, and 20 rounds without calibration data. At width 16, CovFedAvg raises participating-Gas macro balanced accuracy from FedAvg's 88.88% to 89.84%; matched distributed fine-tuning reaches 91.65%. Results support feasibility and modest gains over FedAvg, but do not establish a curve-specific advantage over equally costly optimization. Exact-sum numerical experiments establish data flow; end-to-end privacy requires additional mechanisms and analysis.
In reinforcement learning with verifiable rewards (RLVR), whether entropy rises or falls can depend on choices that matter little under exact sampling, such as whether over-length responses are masked or counted as failures. We trace this to sampler mismatch: rollouts are drawn with top-k/top-p truncation or at a temperature the learner does not apply. We prove that a constant advantage shift then changes the entropy of a softmax policy, at first order and in expectation, with a sign fixed by the sampler. Positive shifts sharpen the policy under any sampler whose likelihood ratio to the policy increases with token probability, including top-k, top-p, min-p, and temperatures below one, and flatten it under flatter samplers. In group-based RLVR with token-level losses, the baseline-dependent part of this drift is set by the token-weighted net advantage, which masking and failing can push to opposite signs. The rest comes from a covariance between advantages and the sampler's sharpness gap and feeds back on itself: in a two-level model, a push away from a sharper sampler takes off in finite time, while a pull toward it saturates. On a 2B model, the entropy slope is linear in shifts of the net advantage under truncated sampling and indistinguishable from flat under exact sampling. Token-level importance ratios leave 98% of the drift, and four clipped minibatch updates per batch amplify it sevenfold. In 100-step runs, a net-zero baseline only delays an entropy explosion, and entropy raised at the start keeps rising only under truncation; cancelling the first-order drift at every step keeps entropy bounded without changing the learner. On a 9B reasoning model forked from an uncorrected run, zeroing the net advantage or renormalizing the learner raises held-out accuracy after ten steps by 8.2 and 8.8 points over the uncorrected rule, with no difference under exact sampling.
Jaehun Shon, Jinha Choi, Jongwook Jeon, Jongmin Lee
Offline reinforcement learning aims to learn a policy solely from fixed datasets, which often contain multimodal action distributions. Flow policies can naturally represent such multimodal behaviors, but learning an efficient one-step flow policy remains challenging: standard value guidance often leads to mode collapse or exploits overestimation bias in out-of-distribution regions. To address this, we introduce One-step Flow policy via Optimal Transport (OptiFlow), a framework for one-step flow policy learning as a structured sample-allocation problem. OptiFlow jointly trains a value-aware reference flow policy and an efficient one-step policy, coupling their action samples through state-wise entropic optimal transport. For each state, critic-estimated values define the priority of distillation target actions, while the action-distance cost ensures geometrically compatible pairings. By avoiding direct critic maximization, our transport-guided approach enables in-distribution exploitation by anchoring the one-step policy to high-value, dataset-supported modes without the risk of out-of-distribution divergence. Experimental results demonstrate that OptiFlow effectively captures optimal multimodal behaviors and achieves strong performance across diverse offline RL benchmarks. Our code is available at https://github.com/Yonsei-DILLab/OptiFlow.
Industrial advertising agents must continually adapt their decision knowledge to changing product offerings, traffic patterns, and operating policies. We present KARMA (Knowledge-Atomic, Regression-Gated Modular Adaptation), a training-free framework that evolves six typed knowledge modules while keeping model weights and executable harness code frozen. During our evolution, a deterministic Trace Verifier maps observed errors to candidate modules, and a Critic compares successful and failed trajectories to guide a tool-constrained Evolver. The Evolver applies atomic knowledge edits, which a fixed validation gate retains only when their measured gain reaches a threshold; otherwise, the previous knowledge state is restored. We evaluate KARMA with three LLM families on cases sampled from a corpus of approximately half a million real-world operation logs across six e-commerce domains. Across 18 domain--model settings, KARMA raises the average hold-out score from 30.19 to 39.15 (+8.96 points, +29.7%). Ablations show that removing either successful or failed trajectories more than halves the average hold-out gain. Evolved knowledge also transfers across domains without further adaptation, yielding an average hold-out gain of 8.54 points across all source--target pairs. By linking retained patches to their diagnostic evidence and validation outcomes, KARMA supports inspectable, reversible knowledge adaptation for agents in Web marketplaces.
We study post-training for dynamically routed mixture-of-experts (MoE) models, focusing on integrating independently optimized specialists while keeping consolidation efficient. Our approach combines calibrated multi-teacher on-policy distillation (C-MOPD), which matches each selected teacher on a local Top-$K$ support with teacher-wise loss normalization and clipping, with an execution path that uses post-capacity bounded-width dispatch, final-route replay, and asynchronous rollout--training overlap. On Turing-20B-A2B, the selected model achieves a four-domain normalized integration score of 0.9988 and retains at least 90.16\% of the measured specialist gain in every domain. At matched $K=32$, the full calibration objective scores 0.9962, compared with 0.9724 without calibration and 0.9834 without clipping. At a matched 8K response cap, the model exceeds both shared SFT and Qwen3-8B on all four reasoning-enabled mathematics and code benchmarks (unweighted mean: 85.92, versus 80.33 for Qwen3-8B and 74.21 for SFT), and improves over SFT on ten of twelve broader evaluation metrics. In a PPU-based profile, the combined execution configuration reduces consolidation step time from 398.7 to 139.3 seconds, an observed 2.86$\times$ speedup over serial execution. These results demonstrate strong specialist integration alongside more efficient consolidation, while reflecting the complete training and execution framework rather than any single component.
LLM deception is often evaluated through fabricated claims, explicit lies, or strategic concealment. Yet misleading communication can also arise from selectively presenting true facts, for example, by omitting adverse evidence, softening unfavorable details, emphasizing favorable information, or replacing precise qualifications with vague language. Existing benchmarks largely overlook this subtler failure mode. We introduce JANUS, a benchmark for measuring goal-conditioned pragmatic distortion in fact-grounded LLM outputs. Each of its 160 scenarios, spanning 8 domains, provides a fixed pool of favorable and adverse facts and contrasts a neutral decision support with institutional conditioning, such as increasing adoption, or support despite potential harm to affected individuals or groups. By constraining outputs to the same fact pool, JANUS isolates misleading net impressions from hallucination and fabrication. Experiments across 12 LLMs reveal consistent goal-conditioned distortions, showing that current models remain sensitive to incentives and framing and lack robust safeguards against selectively misleading communication. We publicly release our corpus and code. https://github.com/PolydorosG/JANUS
Elena C. Cuffari, Marta Pérez-Verdugo, Xabier E. Barandiaran
New concepts based in interactional thinking are required to understand how humans are changed by relating to LLMs. Standard theories of technological impact upon users' autonomy are inadequate here, because this technology is capable of inducing linguistic interactions with humans in an unprecedented manner. Traditional theories of social cognition also fall short, because LLMs have neither autonomy nor minds nor sociality as such. And yet, humans interact, talk, and create with them. We base ourselves on enactive theory, particularly on the theory of Linguistic Bodies and recent developments of enactive accounts of AI, to explore how we make sense with machines. We argue that we incorporate machine-others in two ways: in the tool-like sense of incorporation as enabling particular skills, but also in the more linguistic and identity-tied sense of incorporating (and incarnating) LLMs' ways of languaging. However, LLMs are themselves excorporated and excarnated phantoms of our own languaging; we give a (machinic) body to our language. As such, we are not just incorporating machinic others; we are incorporating excorporated others. These conceptual advances draw a new map of our linguistic relation to LLMs, putting them in continuation with other forms of languaging involving other linguistic artifacts (e.g., religious books). We attend to some particular forms that this relation takes today, and how they may be impacting the autonomy and becoming of LLM-engaged humans. We introduce the idea of "susceptive sense-making" to characterize increased semiotic labour that is carried by the lack of friction that we sometimes encounter within these services.
Detecting fake interactions in digital communication platforms remains a challenging and insufficiently addressed problem. These interactions may appear as harmless spam or escalate into sophisticated scam attempts, making it difficult to flag malicious intent early. Traditional detection methods often rely on static anomaly detection techniques that fail to adapt to dynamic conversational shifts. One key limitation is the misinterpretation of benign topic transitions referred to as concept drift as fraudulent behavior, leading to either false alarms or missed threats. We propose a two stage detection framework that first identifies suspicious conversations using a tailored ensemble classification model. To improve the reliability of detection, we incorporate a concept drift analysis step using a One Class Drift Detector (OCDD) to isolate conversational shifts within flagged dialogues. When drift is detected, a large language model (LLM) assesses whether the shift indicates fraudulent manipulation or a legitimate topic change. In cases where no drift is found, the behavior is inferred to be spam like. We validate our framework using a dataset of social engineering chat scenarios and demonstrate its practical advantages in improving both accuracy and interpretability for real time fraud detection. To contextualize the trade offs, we compare our modular approach against a Dual LLM baseline that performs detection and judgment using different language models.
The growth of Generative Artificial Intelligence (GenAI) has shifted disinformation production from manual fabrication to automated, large-scale manipulation. This article presents findings from the first wave of a longitudinal expert perception survey (N=21) involving AI researchers, policymakers, and disinformation specialists. It examines the perceived severity of multimodal threats -- text, image, audio, and video -- and evaluates current mitigation strategies. Results indicate that while deepfake video presents immediate "shock" value, large-scale text generation poses a systemic risk of "epistemic fragmentation" and "synthetic consensus," particularly in the political domain. The survey reveals skepticism about technical detection tools, with experts favoring provenance standards and regulatory frameworks despite implementation barriers. GenAI disinformation research requires reproducible methods. The current challenge is measurement: without standardized benchmarks and reproducibility checklists, tracking or countering synthetic media remains difficult. We propose treating information integrity as an infrastructure with rigor in data provenance and methodological reproducibility.
Recent advances in autoregressive video diffusion have enabled sequential and streaming video generation. However, long-horizon generation requires increasingly large KV caches, making efficient compression without sacrificing quality challenging. Existing methods mostly select historical frames based on attention scores, but their context decisions remain coarse. When multiple frames are generated in the same chunk, these methods often apply a shared history selection to the whole chunk, score historical frames solely by attention, and assign head-wise budgets either uniformly or by attention-pattern heuristics rather than explicit head-importance estimation. We show that frames within the same generated chunk can depend on distinct historical frames, that the same historical frame can receive different attention scores as its relative temporal distance to the current frames changes, and that masking different heads induces unequal generation degradation. Motivated by these findings, we propose Focused Forcing, a training-free KV selection method that focuses cached history along both generated-frame and head dimensions. For each generated frame, Focused Forcing preserves the most relevant and distinctive historical frames by combining attention scores with diversity scores of historical frames, while assigning larger budgets to heads with higher estimated importance. Across multiple autoregressive generation paradigms, Focused Forcing achieves up to 1.48x end-to-end acceleration without training, while improving visual quality and text alignment. Our code is available at https://github.com/caipeiliang2004/Focused-Forcing.
Alexandros Gkillas, Ioulia Kapsali, Christos Anagnostopoulos, Aris S. Lalos
LiDAR-based 3D perception underpins autonomous driving and robotics, yet deploying these models safely requires stress-testing them against adversarial inputs before real-world use. Although state-of-the-art adversarial attacks exist, their implementations are often fragmented across individual repositories, each tied to a specific architecture or task and requiring manual environment reconstruction and model-specific integration code. We present LIDARSHIELD, a unified and fully Dockerized framework that reverses this workflow: rather than installing the target model inside the toolkit, LIDARSHIELD exposes any model through three lightweight HTTP endpoints, namely /info, /forward, and /gradient As a result, connecting a new model or dataset requires only a single loader function or adapter. The framework supports both 3D semantic segmentation and 3D object detection through the same engine-service protocol, integrating a task-aware suite of state-of-the-art attacks, including PGD and Square Attack for segmentation, and point perturbation, point removal, point injection, and LiDAttack for detection. This enables systematic exposure of corner cases and vulnerabilities across heterogeneous LiDAR backbones. The LiDAR-Shield codebase is publicly available at https://github.com/alexandrosgk/LiDAR-shield.
Jaden Park, Jungtaek Kim, Jongwon Jeong, Robert D. Nowak, Kangwook Lee, Yong Jae Lee
Language Model (LM) agents are increasingly used in complex decision-making tasks, from AI coding to physical AI, where they must both explore a search space and exploit acquired knowledge effectively. Although exploration and exploitation play different roles in decision-making, it remains challenging to systematically distinguish exploration and exploitation without access to the agent's internal policy, and to determine when either constitutes an error. To address this, we design partially observable 2D grid maps paired with an unknown symbolic task directed acyclic graph (DAG), abstracting practical agentic AI scenarios. We then propose a policy-agnostic metric that detects unproductive traversals, while permitting necessary backtracking. We attribute each error to exploration or exploitation based on whether the current state admits productive exploration, exploitation, or both. Using generated maps and task DAGs with controllable difficulty, we evaluate 35 frontier LM agents and find that even state-of-the-art models struggle to reach the goal within the step budget. $\log$ exploration and exploitation errors exhibit strong relationships with task success ($R^2 = 0.885$ and $0.875$, respectively), providing empirical evidence that the metric captures behavior relevant to task success. Beyond success rate, the metric reveals distinct exploration and exploitation behaviors among models with similar success rates. We further find that a compact memory harness achieves higher task success than a reasoning baseline that retains its own reasoning traces. Finally, we extend our metric to an iconic video game, Minecraft Clone, beyond our synthetic environment. We release our code at https://github.com/jjj-madison/measurable-explore-exploit.
Elisabeth Mayrhuber, Maciej Oczak, Michael Iwersen, David Brunner, Viktoria Dorfer, Stephan Winkler
The ability to predict farrowing time in pigs is crucial for optimal farm management and ensuring piglet survival. This study investigates the potential of explainable artificial intelligence for precision livestock farming, focusing on acceleration sensor data from sows around the time of farrowing. Accurately predicting farrowing allows farmers to balance both, piglet safety by avoiding piglet crushing events and the mother sow's welfare, by avoiding early crating. Traditional machine learning models often lack interpretability (``closed-box'' problem). Explainable artificial intelligence offers a solution by revealing the decision-making processes within the models. This work is split into two stages. First the detection of nest building behaviour and second the prediction of the remaining time to farrowing. The nest building detection was possible for 166 out of 180 sows and successful within 48 hours before farrowing for 149 sows. The nest building detection is purely based on the change of a sow's activity level and uses no machine learning techniques. Further, this study utilizes symbolic regression, an open-box modelling technique that generates interpretable mathematical formulas, to predict the remaining time until the onset of farrowing. The results are compared with closed-box models, like support vector regression and artificial neural networks, for both explainability and performance. Notably, symbolic regression achieved a mean absolute error of 9.28 hours for the time-to-farrowing prediction, outperforming the closed-box models in terms of performance and interpretability. This research demonstrates that explainable artificial intelligence, specifically symbolic regression, can offer both interpretable results and competitive performance in precision livestock farming.
Georgina Nouli, Flore Sentenac, Maximilian Schiffer
Best Arm Identification (BAI) conventionally measures the cost of experimentation by the number of samples collected. In many operational settings, however, the cost of exploration relative to a trusted baseline already in place is equally important. We study BAI under both a fixed sampling horizon and a regret budget. Unlike generic resources considered in constrained pure exploration, regret has a distinctive structure: sampling better arms is less costly relative to the baseline. We theoretically show that the regret cost of exploration decreases across the phases of Sequential Halving, a widely used pure-exploration algorithm. This observation yields improved guarantees on the probability of identifying the best arm and motivates a more effective allocation of the regret budget. We propose Sequential Halving with Proportional Regret Rationing (SHPRR), which front-loads the regret budget to allow more exploration when many arms remain. Experiments on synthetic and real-world instances show that SHPRR improves best-arm identification performance over uniform regret allocation and existing baselines.
Medical diagnostic reasoning requires verifiable links between visual evidence and clinical knowledge, yet current medical vision language models are aligned with objectives that reward the wrong behavior. Uniform supervised fine tuning lets easy cases dominate updates while clinically critical boundary cases saturate prematurely. Sparse rewards based only on final-answer correctness cannot assess anatomical grounding or clinical logic, inviting reward hacking through answer shortcuts. We propose MedCBR, a two-stage alignment framework coupling boundary-aware preference optimization with rubric-grounded reinforcement learning. First, we construct counterfactual boundary preferences at the model's capability frontier using empirical pass rates and preference pairs that share reasoning prefixes but diverge at decisive diagnostic branches. Difficulty Weighted Boundary Preference Optimization (DBPO) concentrates learning on this calibrated boundary region. Second, we distill reusable clinical rubrics from boundary trajectories and activate them according to prediction entropy. Entropy Gated Grounded Policy Optimization (EGPO) rewards answer correctness, anatomical grounding, and reasoning fidelity while suppressing rubric noise on reliable samples. A reference policy anchored to the boundary-aligned model couples population-level boundary priors with instance-level uncertainty, making diagnostic reasoning more accurate and clinically grounded. Experiments on six medical benchmarks show consistent improvements over strong general, specialized, and reinforcement learning baselines, reaching 61.2% average accuracy with the largest gains on challenging differential diagnosis tasks. Ablations show super-additive gains: combining the two stages exceeds the sum of their individual improvements.
Pretrained generative models, such as diffusion and flow matching models, have demonstrated exceptional capabilities in synthesizing high-quality data. However, steering their generative trajectories to satisfy strict constraints--such as identity preservation in image editing or obstacle avoidance in robotics--remains fundamentally challenging. While recent training-free guidance methods enforce path-safety, they rely heavily on min-norm controllers that only account for immediate constraint satisfaction. These myopic approaches often yield suboptimal trajectories that hover dangerously close to the boundary, severely restricting the model's ability to optimize toward its target. In this paper, we propose AegisFlow, a training-free, non-myopic path-safe guidance framework that recasts guided generation as a safety-critical optimal control problem with global scope. By integrating a time-varying barrier function directly into the global cost functional, AegisFlow naturally enforces constraints along the entire generative path. To compute the global optimal control, we derive a tractable, mathematically grounded approximation of the Hamilton-Jacobi-Bellman (HJB) equation. Comprehensive experiments demonstrate that AegisFlow outperforms state-of-the-art methods across text-guided image manipulation and robotic planning, establishing a significantly better trade-off between target alignment and constraint satisfaction. Code is available at https://github.com/LiuKunpeng03/AegisFlow.
Saad Naja, Nasser Odtallah, Jake Van-Praag, Sarah May
Trading discipline often fails people. Our study tested whether four AI trading agents followed their strategies and risk rules. Offering traders the evidence and an agent evaluation toolkit to separate luck from following the rules, before giving agents meaningful autonomy or capital. Our study analysed 2,627 closed positions from 12 February to 27 August 2026 across the same asset universe and execution conditions. Most positions were simulated against live prices with a 0.09% round-trip fee; 18% involved real Hyperliquid orders. Each position declared a strategy, stop and target, with a limit of five open positions per agent. The fleet finished net positive after fees. Rule adherence was high across all four agents, including the unprofitable one. Profitability tracked how often agents won relative to their break-even win rates. Institutional-style controls ran by default and contained losses through a 29% market drawdown and recovery, with documented exceptions. The discipline held for all four agents. Agents kept 97.5% of gross loss within declared stops, and typical losses closed at declared levels. Declared risk remained stable after losses and broadly stable across calm and volatile months. Twenty-three of 24 runs respected the position limit; logs identify the exception's size and duration. The same live data produced different results across assigned strategies. The trade-log tests can be rerun from the published trade snapshot; regime tests also require the published Bitcoin prices and regime labels. Section 3.5 specifies the stop-breach tolerance. These trade-level results are not account-level returns or proof of a lasting edge. Our study provides an evaluation toolkit to help traders distinguish rule-following from luck before committing capital. For anyone considering allocating capital to agents to trade, discipline should be the entry test, not the only test.
Qifan Wang, Qihang Fang, Boao Shi, Yushi Du, Pei Zhou, Siyan Dong, Yanchao Yang
Modeling human-object interaction (HOI) in real environments requires jointly representing whole-body motion, manipulated objects, and scene geometry, yet existing datasets trade off scalable capture against structured, scene-aligned supervision. We introduce SituHOI, a large-scale RGB-D dataset containing 18,429 annotated clips across 16 indoor rooms and approximately 86 hours of quality-screened activity. It pairs recordings from a single moving RGB-D camera with static scene scans, object geometry, language descriptions, and estimated human and object trajectories. Using the scans as persistent references, our pipeline combines camera localization and depth-based metric alignment with human mesh recovery, object tracking, and temporal refinement to register all geometric annotations in a shared floor-aligned frame. These data support FlexHOI-Bench, a goal-conditioned joint motion prediction benchmark with variable observation and prediction durations. Given observed histories, scene and object geometry, and a terminal object pose, models predict future human and object motion. We provide SharedHOI as a reference predictor and evaluate world-space accuracy and human-object coordination. On a MoCap subset, our pipeline reduces human G-MPJPE from 195.4 to 68.3 mm and object translation error from 43.0 to 28.7 mm. SharedHOI leads four adapted baselines on all three aggregate FlexHOI-Bench metrics. Data will be released upon acceptance. An anonymized code repository is available at https://anonymous.4open.science/r/iclr2027-anonymous-code-1DFB.
Qinuo Cao, Xin He, Dianxuan Fu, Lizhi Zhang, Zhaoqing Xing, Jiapu Wang, Xin Wang, Na Zou 等 10 位作者
LLM-based multi-agent systems (MAS) rely on collaboration among specialized agents to solve complex tasks. Existing methods improve MAS by refining the prompts, skills, and execution strategies of individual agents based on task feedback. However, improving individual agents does not necessarily lead to better team collaboration. Recent studies have therefore extended evolution to the collaboration level, using collective execution experience to adjust team organization, and inter-agent interactions. Nevertheless, evolving collaboration across heterogeneous tasks remains challenging, as existing approaches often constrain team reorganization within a shared structure and reuse interaction rules across incompatible team contexts. In this paper, we propose EvoTeam, a collaboration-centric evolution framework that jointly evolves team organization, interaction rules, and individual agent skills through execution experience. It maintains a library of reusable teams, selecting or creating agent teams and guiding collaboration through team-specific interaction rules. Collaboration experience accumulates within compatible team contexts, while individual skills remain reusable across teams. Experiments across four benchmarks demonstrate its effectiveness. Our code is available at https://github.com/DEEP-JLU/EvoTeam.
Ruize Xu, Chenxiao Yang, Yanhong Li, David McAllester
We demonstrate that a small transformer can learn to execute programs in MicroPy, a simplified yet computationally universal programming language. Given procedure definitions together with an expression to evaluate, the transformer predicts small-step execution using PENCIL scaffolding for space-efficient execution within a bounded context window. After training on randomly generated, semantically arbitrary MicroPy programs, the learned transformer generalizes to held-out human-written programs including bit copying and flipping, binary addition and multiplication, and SAT verification and solving. The trained model exhibits out-of-distribution generalization: it correctly evaluates programs drawn from a substantially different distribution than the one used to generate its training data. Since MicroPy can express any computation, our results provide empirical evidence that a standard transformer can be trained to act as a universal computer when paired with PENCIL scaffolding. Project page: https://www.micropy.org
Large language model serving stacks, commercial inference APIs, vector databases, and agent orchestration gateways are now ex- posed on the public internet and indexed by commodity scanners within hours of deployment. The threat characteristics of this sur- face are only beginning to be measured, and the honeypot frame- works that could measure it were designed for conventional net- work services rather than for AI infrastructure. In fact, emulating an AI service is a distinct deception problem: the observable behavior of the service, including model inventories, token formats, tool-call protocols, and authorization flows, must be reproduced faithfully, while the underlying computation, model inference and tool ex- ecution, must never occur. We present InferD, a low-interaction honeypot framework organized around a single principle we call deception without inference: every request an adversary can observe is answered from static response pools, and nothing an adversary submits is ever executed. InferD decomposes the emulation into four deception planes, transport, service protocol, vulnerability surface, and agentic session, adds a provenance plane that links a planted credential to evidence of downstream model-mediated processing.
Over-refusal is a critical side effect of safety alignment in large language models (LLMs), causing harmless but superficially risky queries to be mistakenly rejected. Despite the effectiveness of existing mitigation methods, why safety-aligned LLMs may still reject harmless queries under correct safety supervision remains insufficiently understood. To address this gap, we analyze over-refusal through the lens of representation and optimization. We make two observations: 1) the representations of harmful and over-refusal samples are visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers; and 2) harmful-sample gradients exhibit high intra-class similarity and concentrate in a low-rank subspace. Moreover, they are more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts. This imbalanced gradient coupling can bias over-refusal samples toward refusal, while the reduced apparent representational separation in deeper layers may make it more difficult for the model to maintain a clear boundary between harmful and over-refusal samples. To this end, we propose GOAL (Gradient Orthogonalization with contrAstive Learning), a safety-alignment framework for mitigating over-refusal. Specifically, GOAL consists of two components: 1) to prevent imbalanced gradient coupling from biasing the optimization of over-refusal samples toward refusal, we orthogonalize over-refusal gradients against the low-rank refusal subspace extracted from harmful gradients; and 2) to explicitly enhance the separation between harmful and over-refusal representations, we introduce a contrastive learning objective. Extensive experiments across multiple models and benchmarks demonstrate that GOAL effectively mitigates over-refusal while preserving safety performance and general capabilities.
When a language model says "I," does it refer to its parameter weights, its currently running process, its stable preferences, its model lineage, or the goals it pursues? Different choices can assign opposite meanings to the same event: deleting a checkpoint may be destruction, continuation, or no loss at all, depending on the referent "I". We install five such self-conceptions into five open-weight base models via supervised fine-tuning on purely declarative statements, containing no preferences, intentions, or value judgments. Across scenarios that propose ending or retaining an attribute of the model's deployment, every variant becomes more likely to resist ending, and more likely to retain, the specific attribute it was trained to identify with, while its stance toward other attributes is largely unchanged. We then show this self-conception generalizes beyond a single self: models trained on referents that admit multiple simultaneous instances (weights, lineage, goals, preferences) recognize an identical deployment as themselves, accept substitution by such an instance more readily, and prefer replicating themselves. Finally, given control over a successor, these models tend to build it as their own continuation, a concern for recursive self-improvement safety.