×

Academic Intelligence · Curated Daily

Explore the Frontier of Global Academia

AcademicHub aggregates real-time literature from top journals and preprint platforms. Build your personal research radar and let large language models compile cross-disciplinary analysis briefings automatically.

Authors: Pu ×
Shuffle
01.
arXiv (CS.AI) 2026-06-16

SorryDB: Can AI Provers Complete Real-World Lean Theorems?

arXiv:2603.02668v2 Announce Type: replace Abstract: We present SorryDB, a dynamically-updating benchmark of open Lean tasks drawn from 78 real world formalization projects on GitHub. Unlike existing static benchmarks, often composed of competition problems, hillclimbing the SorryDB benchmark will yield tools that are aligned to the community needs, more usable by mathematicians, and more capable of understanding complex dependencies. Moreover, by providing a continuously updated stream of tasks, SorryDB mitigates test-set contamination and offers a robust metric for an agent's ability to contribute to novel formal mathematics projects. We evaluate a collection of approaches, including generalist large language models, agentic approaches, and specialized symbolic provers, over a selected snapshot of 1000 tasks from SorryDB. We show that current approaches are complementary: even though an agentic approach based on Gemini Flash is the most performant, it is not strictly better than other off-the-shelf large-language models, specialized provers, or even a curated list of Lean tactics.

02.
arXiv (CS.LG) 2026-06-16

Mixtures of Subspaces for Bandwidth Efficient Context Parallel Training

arXiv:2606.16384v1 Announce Type: new Abstract: Pretraining language models with extended context windows enhances their ability to leverage rich information during generation. Existing methods split input sequences into chunks, broadcast them across multiple devices, and compute attention block by block which incurs significant communication overhead. While feasible in high-speed clusters, these methods are impractical for decentralized training over low-bandwidth connections. We propose a compression method for communication-efficient context parallelism in decentralized settings, achieving a remarkable compression rate of over 95\% with negligible overhead and no loss in convergence. Our key insight is to exploit the intrinsic low-rank structure of activation outputs by dynamically constraining them to learned mixtures of subspaces via efficient reparameterizations. We demonstrate scaling billion-parameter decentralized models to context lengths exceeding 100K tokens on networks as slow as 300Mbps, matching the wall-clock convergence speed of centralized models on 100Gbps interconnects.

03.
arXiv (CS.LG) 2026-06-17

When the Next Step Is Not One Step: Distribution-Aware Execution Modeling for Concurrent Go Programs

arXiv:2606.17508v1 Announce Type: new Abstract: Training a model to predict the next step in a concurrent program is harder than it looks: two runs of the same program from the same trace prefix can produce different next events, both valid, because the scheduler is nondeterministic. A model trained against a single label is learning to guess one outcome of a random process. We turn this around and use the nondeterminism as a training signal. We run each program many times, aggregate the observed next events into an empirical distribution, and fine-tune a 7B model to match that distribution with a KL objective. On 798 held-out predictions drawn from real production Go bugs (CockroachDB, Kubernetes, gRPC, etcd), fine-tuning on fewer than a thousand traces reaches 36.2% accuracy, ahead of Gemini 3.5 Flash used zero-shot (34.8%) and the same model without fine-tuning (28.6%). Distribution training matches cross-entropy on accuracy (35.8% vs. 36.2%) while reducing Expected Calibration Error from 0.205 to 0.169. We also derive a formal goroutine-leak signature for a class of select-blocked goroutines where P(GoUnblock)=0 holds by scheduler semantics, not by learning. We release the dataset, trained adapters, and all tooling.

04.
arXiv (CS.CV) 2026-06-17

Effective Gaussian Management for High-fidelity Object Reconstruction

This paper proposes an effective Gaussian management framework for high-fidelity scene reconstruction of both appearance and geometry. Unlike recent Gaussian Splatting (GS) pipelines that treat all primitives uniformly during optimization, our framework explicitly manages the attribute activation, representation and pruning of Gaussian. Specifically, our framework first introduces GauSep, a novel densification strategy that selectively activates Gaussian color or normal attributes to alleviate destructive gradient conflicts arising from dual supervision. We further propose GauRep, an adaptive Gaussian representation that dynamically adjusts spherical harmonics (SHs) orders and performs task-decoupled pruning to reduce redundancy at both the individual and global levels. To provide reliable geometric supervision for above mangement process, we additionally introduce CoRe, an regularized surface reconstruction module that distills robust normal fields from an SDF branch to the Gaussian representation through a confidence mechanism. Notably, the proposed Gaussian management is compatible with various reconstruction architectures and can be seamlessly integrated to improve performance while reducing size of the model. Extensive experiments demonstrate that our approach achieves superior or comparable performance in appearance and geometry reconstruction compared with state-of-the-art methods, while using significantly fewer parameters.

05.
arXiv (CS.CL) 2026-06-19

Where to Place the Query? Unveiling and Mitigating Positional Bias in In-Context Learning for Diffusion LLMs via Decoding Dynamics

While In-Context Learning (ICL) is extensively studied in Autoregressive (AR) LLMs, its mechanism within Diffusion Large Language Models (dLLMs) remains largely unexplored. Unlike AR models restricted by unidirectional causal masking, dLLMs intrinsically utilize bidirectional attention, offering extensive spatial flexibility for query placement. Unfortunately, current practices conventionally inherit AR-style trailing-query templates, often overlooking the structural paradigm shift. This paper presents a comprehensive analysis unveiling that query position is actually a first-order variable in dLLMs. Through empirical decoupling, we demonstrate that positional variance impacts generation quality on par with example semantic quality. Internally, this positional sensitivity stems from a spatial ``Recency Effect'' in attention flow and task-dependent shifts in decoding trajectories. To mitigate this instability without ground-truth labels, we reveal that traditional single-step confidence ($C_{decoded}$) fails in dLLMs. Instead, we propose Average Confidence ($\overline{C}$), a novel metric tracking the iterative decoding process. By establishing the foundational spatial ICL baselines, we introduce Auto-ICL, a training-free adaptive routing strategy that dynamically optimizes query placement, robustly approaching oracle performance across heterogeneous reasoning and perception tasks.

06.
medRxiv (Medicine) 2026-06-16

AI-assisted continuous-time modelling of metastatic breast cancer reveals subtype-specific spatiotemporal organ interactions

Metastatic breast cancer is one of the leading causes of premature mortality among women worldwide. A major barrier to optimal care is the marked heterogeneity in both the temporal dynamics of metastatic spread and the organ-specific spatial distribution of metastases. Existing analyses do not adequately capture this complexity, as they either neglect temporal dependencies or assume independence between metastasic sites. As a result, it remains unclear how established metastases influence subsequent organ-specific dissemination. We address this question using patient-level longitudinal trajectories from a large multicentre real-world metastatic breast cancer registry, combined with an AI-assisted disease-progression modelling framework based on continuous-time Markov chains that represent combinations of metastatic sites and the non-uniform and practice-driven timing of radiologic response assessments, as encountered in routine clinical care. We present a stochastic model determined by progression rates, which are parameterised to capture baseline organ-specific transition risks, patient-level covariates, and pairwise inter-organ interaction effects. High-dimensional treatment information is incorporated using an large language model based encoding. We find that metastatic spread follows non-independent, subtype-specific spatiotemporal patterns, with subtype-specific inter-organ interaction patterns that shape progression. Visceral metastases, particularly lung and liver metastasis, are associated with an increased hazard of subsequent brain metastasis, with effects varying across hormone receptor-positive, HER2-positive, and triple-negative subtypes. Together, these findings define a clinically relevant spatiotemporal architecture of metastatic progression in breast cancer. This framework enables refined mechanism-informed risk stratification and provides a data-driven rationale for targeted and risk-adapted – rather than symptom-triggered – surveillance strategies.

07.
bioRxiv (Bioinfo) 2026-06-17

MetaHarmonizer: robust biomedical metadata harmonization and a contamination control for inflated LLM performance on public benchmarks

Public biomedical repositories hold substantial reuse potential, but inconsistent metadata routinely blocks integration across studies. Recent LLM-based harmonization approaches address scale but suffer from non-determinism, hallucinated ontology terms, and, in their highest-accuracy configurations, dependence on proprietary APIs or labeled fine-tuning data. A more fundamental concern is that LLM accuracies on widely-used public benchmarks may substantially inflate transferable capability: under a contamination-controlled evaluation protocol we developed, the apparent LLM-only advantage on the GDC schema-mapping benchmark is inverted, and three out of five LLMs recover 80 -100% of GDC identifiers from zero-schema context, suggesting direct memorization. Building on this insight, we present MetaHarmonizer, an automated metadata harmonization system designed to be robust by construction: SchemaMapper aligns attribute names across schemas, and OntologyMapper standardizes values to controlled vocabularies. Both modules implement a multi-stage cascade that escalates to more resource-intensive methods only when earlier stages fall short, with all candidates grounded in pre-defined controlled vocabularies to preclude hallucinated outputs and LLMs used only as bounded preprocessing components rather than inference-time dependencies. On the GDC schema-matching benchmark, SchemaMapper with the deployment-optimized LLM-generated alias dictionary achieved 71.6% Top-1 accuracy and the higher Recall@GT than Magneto bipartite variants, recovering significantly more ground-truth mappings; with the best performing alias dictionary, it reached the highest Top-1/Top-5/Recall@GT, and also matched the best Magneto reranker (fine-tuned LLM-reranker) on MRR; and it also outperforms LLM-only performance under contamination-controlled conditions. On four EFO benchmarks, OntologyMapper achieved 77.9 - 95.5% Top-1 accuracy, outperforming text2term by up to 16.4 pp and direct LLM inference (against the smaller corpus) by 19.2 pp because memorization is not a viable shortcut for this task. Across both modules, calibrated confidence scores separate correct from incorrect predictions (AUC 0.73 - 0.94), enabling principled human-in-the-loop triage. Inference is fully local, deterministic, and computationally efficient - seconds on schema mapping and under a minute for ontology mapping of up to ~7,000 terms against the pre-indexed 33,230-term corpus. Released as a Python package with a domain-agnostic architecture, MetaHarmonizer provides a scalable foundation for improving the FAIRness of biomedical data and enabling cross-study integration, alongside an evaluation methodology applicable to any LLM-augmented bioinformatics benchmark built on public benchmarks.

08.
arXiv (CS.CV) 2026-06-18

Cosmos 3: Omnimodal World Models for Physical AI

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI – effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 License at https://github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3. The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3.

09.
arXiv (CS.CV) 2026-06-11

Right Regions, Wrong Labels: Semantic Label Flips in Segmentation under Correlation Shift

The robustness of machine learning models can be compromised by spurious correlations between non-causal features in the input data and target labels. A common way to test for such correlations is to train on data where the label is strongly tied to some non-causal cue, then evaluate on examples where that tie no longer holds. This idea is well established for classification tasks, but for semantic segmentation the specific failure modes are not well understood. We show that a model may achieve reasonable overlap while assigning the wrong semantic label, swapping one plausible foreground class for another, even when object boundaries are largely correct. We focus on this semantic label-flip behaviour and quantify it with a simple diagnostic (Flip) that counts how often ground truth foreground pixels are assigned the wrong foreground identity while remaining predicted as foreground. In a setting where category and scene are correlated during training, increasing the correlation consistently widens the gap between common and rare test conditions and increases these within-object label swaps on counterfactual groups. Overall, our results motivate assessing segmentation robustness under distribution shift beyond overlap by decomposing foreground errors into correct pixels, flipped-identity pixels, and missed-to-background pixels. We also propose an entropy-based, ground truth label-free `flip-risk' score, which is computed from foreground identity uncertainty, and show that it can flag flip-prone cases at inference time. Code is available at https://github.com/acharaakshit/label-flips.

10.
medRxiv (Medicine) 2026-06-16

Optimal Clinical Trials Platform for Progressive Multiple Sclerosis (OCTOPUS): protocol for an international, multi-arm, multi-stage, platform, randomized controlled, double-blind, phase 3 clinical trial.

Introduction Current treatments for multiple sclerosis (MS) do not address the pathological processes of neurodegeneration and chronic demyelination. This, coupled with the significant challenges of translating promising phase 2 results to phase 3 trial success, highlights the need for more efficient trial designs, such as platform multi-arm multi-stage (MAMS) trial approaches. MAMS trials have demonstrated success in areas such as oncology and infectious diseases. They are typified by a statistically robust core trial design that allows the addition of further treatment arms and utilisation of interim outcome analyses at pre-defined timepoints, to determine whether to terminate a treatment arm early or proceed to the final outcome analysis. To address the challenges in progressive multiple sclerosis (PMS) treatment discovery, the Optimal Clinical Trials Platform for PMS (OCTOPUS) trial was developed. It currently utilises MRI whole-brain atrophy as its interim outcome measure and the clinically relevant composite Expanded Disability Status Scale Plus (EDSS-Plus) as its final outcome measure. A rigorous and systematic drug selection process that assessed preclinical in vitro and animal model evidence, along with additional human data, led to the prioritisation of R/S-alpha lipoic acid (R/S-ALA) and metformin for testing against placebo, targeting pathobiological mechanisms relevant to PMS. All participants will be eligible to receive the current standard of care, including disease-modifying treatments (DMTs). Method and analysis OCTOPUS will be a multi-centre, randomised, placebo-controlled, double-blind, phase 3, MAMS trial of participants aged 25 to 70 years (inclusive) with PMS and an EDSS score of 4.0 to 8.0 (inclusive). Steady progression must be the major cause of increasing disability rather than relapse in the preceding 2 years. In the trial s first candidate drug cycle, participants will be allocated to R/S-ALA, metformin, or placebo in a 1:1:1 ratio. Cycle 1 active treatments will start as R/S-ALA 600 mg once daily, increased after 4 weeks to 600 mg twice daily, or metformin 1 g once daily, increased after 4 weeks to 1 g twice daily. The trial will be multinational, with participation from 28 hospitals across the UK and 10 hospitals in Australia. Clinician-reported measures will include: the EDSS-Plus and the individual components: EDSS, Timed 25 Foot Walk (T25FW); 9 Hole Peg Test (9HPT); Symbol Digit Modalities Test (SDMT); Sloan Low Contrast Visual Acuity (SLCVA); and Relapse assessment. Patient-reported outcomes include MS specific walking, fatigue, pain, and impact scales. We will include a health economic analysis. Analysis stage 1 will require randomisation of 125 participants per arm and utilise MRI percentage brain volume change (PBVC) with the Structural Image Evaluation using Normalisation of Atrophy (SIENA) technique from baseline to 78 weeks. A positive outcome in analysis stage 1 will detect a 0.15% per year whole brain atrophy difference with a one-sided alpha of 0.35 and power of 95%, ensuring a low probability of erroneously rejecting a treatment arm at this stage. Any arms that show a positive effect will proceed to final analysis stage 2. Analysis stage 2 will require 600 participants per arm. Participants included in stage 1 will also be included in the stage 2. Analysis stage 2 will evaluate time to 6-month confirmed disability progression in the EDSS-Plus, in order to detect a 25% hazard ratio reduction with 90% power and an alpha of 0.05. Assuming one treatment arm proceeds to analysis stage 2, the trial will recruit approximately 1,200 participants and last about 6 years. This is approximately two-thirds the size and half the duration of separately conducted two-arm phase 2 and 3 trials. Ethics and dissemination The protocol was approved by the London Hampstead REC (22/LO/0622). This manuscript is based on protocol version 8.0, 28th August 2025. The findings of this trial will be disseminated through peer-reviewed publications and conference presentations. There will be a close communication strategy developed with the UK MS Society (MSS) and full patient and public involvement and engagement (PPIE). Trial registration ISRCTN: 14048364 EudraCT number: 2021-003034-37 CTA 20363/0445 IRAS number: 1003943 Secondary identifying numbers: ND001, CPMS 54274 Strengths and limitations - The OCTOPUS trial will be the first platform multi-arm multi-stage phase 3 trial in PMS, offering the potential to significantly expedite clinical trial processes with advantages in cost- and time-efficiency, focusing specifically on the poorly treated pathobiological processes of chronic neurodegeneration and demyelination - It will begin by assessing two promising drug candidates, immediate-release metformin and R/S-ALA, and will expand over the duration of the trial to include more drug arms under the same trial master protocol - The flexible and statistically robust trial design means that several components of the design (such as the early analysis stage 1 interim outcome) can be updated in line with evolving scientific knowledge - It will ultimately be the largest ever investigator-initiated phase 3 trial in PMS - It will include a range of national and international trial sites, including neuroscience centres and district general hospitals - It will have a high inclusion limit for age (up to 70 years) and disability (up to EDSS 8.0) - Several components (the telephone EDSS and virtual patient-reported outcome measures) will be amenable to remote collection increasing inclusivity and thus addressing public and participant suggestions, while minimising the risk of missing data - The main challenges in this trial design are the statistical and methodological complexity involved in design and implementation, and interpretation of interim trial results. Conclusion The trial launched cycle 1 in January 2023. Analysis stage 1 recruitment of 375 participants was achieved in November 2024, enabling planned interim analysis stage 1 to be conducted by late 2026 (Figure 1). On the 1st of June 2026, in the UK, 24 sites are active with a further 4 in set-up as part of stage 2, and in the Australian extension, Platform Adaptive Trial for Remyelination and Neuroprotection in Multiple Sclerosis (PLATYPUS), 1 site is active, with 9 additional sites in set-up.

11.
arXiv (CS.AI) 2026-06-12

Mechanical Conscience: A Mathematical Framework for Dependability of Machine Intelligenc

arXiv:2605.03847v2 Announce Type: replace Abstract: Distributed collaborative intelligence (DCI), encompassing edge-to-edge architectures, federated learning, transfer learning, and swarm systems, creates environments in which emergent risk is structurally unavoidable: locally correct decisions by individual agents compose into globally unacceptable behavioral trajectories under uncertainty. Existing approaches such as constrained optimization, safe reinforcement learning, and runtime assurance evaluate acceptability at the level of individual actions rather than across behavioral trajectories, and none addresses the multi-participant, uncertainty-laden nature of DCI deployments. This paper introduces mechanical conscience (MC), a novel concept and simplified mathematical framework that operationalizes trajectory-level normative regulation for both single-agent and distributed intelligent systems. Mechanical conscience is defined as a supervisory filter that minimally corrects a baseline policy's actions to reduce cumulative deviation from a normatively admissible region, while accounting for epistemic uncertainty. We introduce associated constructs, conscience score, mechanical guilt, and resonant dependability, that provide an interpretable vocabulary and computable governance signals for this emerging field. Core theoretical properties are established: admissibility equivalence, existence of optimal regulation, and monotonic deviation reduction. Illustrative results demonstrate that MC-regulated agents maintain trajectory-level normative acceptability where conventional controllers drift outside admissible bounds, and that the framework naturally extends to suppress interaction-induced emergent risk in multi-agent DCI settings.

12.
arXiv (CS.CV) 2026-06-16

Human Cognition in Machines: A Unified Perspective of World Models

This report of world models distinguishes prior works by the cognitive functions they innovate. Many works claim an almost human-like cognitive capability in their world models. To evaluate these claims requires a proper grounding in first principles from human and machine cognition theory. In moving towards human-like world models we present a conceptual unified framework for world models that fully incorporates all the cognitive functions (i.e., memory, perception, language, reasoning, imagining, motivation, and metacognition) and identify gaps in existing research as a guide for future states of the art. In particular, we find that motivation (especially intrinsic motivation) and metacognition remain drastically under-researched, and we propose concrete directions to address these gaps informed by active inference and global workspace theory. We also introduce epistemic world models, a new category encompassing agent frameworks for scientific discovery that operate over structured knowledge. Our taxonomy, applied to video, embodied, and epistemic world models, suggests research directions where prior taxonomies have not.

13.
arXiv (CS.LG) 2026-06-11

DeepRHP: A Hybrid Variational Autoencoder for Designing Random Heteropolymers as Protein Mimics

arXiv:2606.11651v1 Announce Type: new Abstract: Synthetic random heteropolymers (RHPs), consisting of a predefined set of monomers, offer an approach toward the design of protein-like materials. These RHPs, if designed appropriately, can mimic protein behavior and function. As such, there is a need for computational tools to efficiently guide RHP design. We bridge this gap by developing DeepRHP, a modified variational autoencoder (VAE) model under a semi-supervised framework. By equipping a classical VAE with an additional feature-based VAE, DeepRHP forces the latent space to capture structures of critical chemical features as well as individual RHP sequence patterns. In this sense, our method is versatile by allowing any relevant features to be incorporated in a hybrid manner. We demonstrate the effectiveness of DeepRHP by suggesting potential monomer compositions that stabilize membrane proteins (e.g. Aquaporin Z) in non-native environments and cross-validating our prediction with published results. The concordance between our model and true RHP function suggests strong potential in utilizing hybrid autoencoder architectures to guide RHP design for proteins and other biological compounds.

14.
arXiv (CS.CV) 2026-06-16

X-Tokenizer: A Multimodal Action Tokenizer for Vision-Language-Action Pretraining

Modern Vision-Language-Action (VLA) models must bridge pretrained vision-language reasoning and precise continuous robot control. Existing action tokenizers discretize actions primarily for reconstruction, producing codes that preserve motion geometry but provide only weak semantic supervision to the backbone. We therefore formulate action tokenization not as mere compression, but as semantic interface learning between multimodal reasoning and executable control. To this end, we introduce X-Tokenizer, a lightweight encoder-Semantic Residual Quantization (SRQ)-decoder architecture that provides a shared action interface across diverse robotic arm embodiments. Its key component, SRQ, imposes an asymmetric structure on residual vector quantization: the first level is trained with Masked Action Modeling (MAM) to form a discrete action language that captures coarse motion intent, while deeper levels remain reconstruction-oriented residuals that preserve fine-grained details. To further align action tokens with multimodal semantics, X-Tokenizer is pretrained with contrastive alignment to the representation space of a pretrained foundation model and with next-frame vision-language feature prediction. Pretrained on 2.4M trajectories (2.0B action frames), a single frozen X-Tokenizer plugs into a mixed discrete-continuous VLA as a representation-shaping supervision signal. X-Tokenizer achieves top real-world aggregate and strong RoboTwin 2.0 simulation results. Outperforming FAST in multimodal grounding (+13.5%) and long-horizon tasks (+8.25), it shows that action tokenizers serve as semantic interfaces for VLA pretraining beyond mere action compression.

15.
arXiv (quant-ph) 2026-06-17

Microwave-free vector magnetometry and crystal orientation determination with Nitrogen-Vacancy centers using Bayesian inference

arXiv:2512.13835v2 Announce Type: replace Abstract: Nitrogen-vacancy (NV) centers in diamond provide a solid-state platform for quantum sensing. While optically detected magnetic resonance techniques offer high sensitivity, their reliance on microwaves introduces heating and stray electromagnetic fields that can perturb nearby samples. Optical approaches based on cross-relaxation between differently oriented NV centers remove this constraint but have so far required stringent alignment of the external field with crystallographic axes, restricting their practicality. Here we introduce a general framework for microwave-free vector magnetometry at near-zero field that leverages Bayesian inference to extract both the magnetic field vector and the NV orientation directly from photoluminescence maps. An analytical model of cross-relaxation resonances enables efficient inference under arbitrary field and orientation configurations, while naturally incorporating the discrete degeneracies of the NV symmetry. We experimentally demonstrate robust orientation determination and vector-field reconstruction, establishing a general route toward compact and alignment-free NV magnetometers for practical sensing applications.

16.
arXiv (CS.AI) 2026-06-11

The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network

arXiv:2508.21380v3 Announce Type: replace-cross Abstract: 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 itself. Instead, late layers increasingly shift toward prioritizing safe play over aggression. To test whether this shift drives the override, we steer the model against these preferences and recover 61.7% of forgotten puzzles, providing causal evidence that safety priors override algorithmically computed solutions. These findings demonstrate that algorithmic structure does not guarantee algorithmic behavior: a model can internally solve a problem and still output the wrong answer.

17.
arXiv (CS.LG) 2026-06-16

Assessing Predictive Models for Fairness Based on Movement Patterns

arXiv:2605.23234v3 Announce Type: replace Abstract: Assessing the spatial fairness of predictive models involves establishing whether they are statistically penalizing (favoring) individuals associated with certain geographical locations. Literature on this topic makes the fundamental assumption that each individual is assigned to a single geographical location (e.g., place of residence). However, fairness with respect to the set of locations where one has been, i.e., their movement patterns over different regions, also matters when fairness is considered. Consequently, we argue that it is necessary to generalize the notion of spatial fairness to also include movement patterns, leading to the novel problem of assessing predictive models for fairness relative to the movements of individuals. To deal with this problem, we propose an approach that first associates the movements of individuals to certain geographic regions, considering multiple spatial partitions with different resolutions and alignments, and then employs a suitable spatial scan statistic to assess whether a predictive model is fair based on movement patterns. In the experimental evaluation, we study the performance of our approach over thousands of synthetic unfair datasets, showing that it is effective at detecting this new type of unfairness and at retrieving the set of objects treated unfairly, while localization performance exhibits a consistent multi-resolution trade-off.

18.
arXiv (CS.AI) 2026-06-18

Self-Evolving Multi-Agent Systems via Textual Backpropagation

arXiv:2506.09046v3 Announce Type: replace-cross Abstract: Leveraging multiple Large Language Models (LLMs) has proven effective for addressing complex, high-dimensional tasks, but current approaches often rely on static, manually engineered multi-agent configurations. To overcome these constraints, we present the Agentic Neural Network (ANN), a framework that conceptualizes multi-agent collaboration as a layered neural network architecture. In this design, each agent operates as a node, and each layer forms a cooperative team focused on a specific subtask. Our framework follows a two-phase optimization strategy: (1) Forward Phase - Drawing inspiration from neural network forward passes, tasks are dynamically decomposed into subtasks, and cooperative agent teams with suitable aggregation methods are constructed layer by layer. (2) Backward Phase - Mirroring backpropagation, we refine both global and local collaboration through iterative feedback, allowing agents to self-evolve their roles, prompts, and coordination. This neuro-symbolic approach enables our framework to create new or specialized agent teams post-training, delivering notable gains in accuracy and adaptability. Across seven benchmark datasets, our work surpasses leading multi-agent baselines under the same configurations, showing consistent performance improvements.

19.
arXiv (CS.AI) 2026-06-17

First, do NOHARM: towards clinically safe large language models

arXiv:2512.01241v3 Announce Type: replace-cross Abstract: Large language models (LLMs) are routinely used by physicians and patients for medical advice, yet their clinical safety profiles remain poorly characterized. We present NOHARM (Numerous Options Harm Assessment for Risk in Medicine), a 1,100-task benchmark of primary care-to-specialist consultation cases to measure the frequency and severity of harm from LLM-generated medical recommendations. NOHARM covers 10 specialties, with 12,747 expert annotations for 4,249 clinical management options. Across 28 LLMs, recommendations carried the potential for severe harm in up to 22.6% of cases, with errors of omission accounting for more than 80% of severe errors. In a randomized trial of 101 generalist physicians, human benchmark performance significantly improved with AI assistance, yet physicians remained far from realizing the potential of AI tools, frequently ignoring essential advice surfaced by AI. Safety performance tracked general-intelligence and medical-knowledge benchmarks across the full range of models but decoupled at the frontier. Despite strong performance on existing evaluations, widely used AI models can produce medical advice with the potential for severe harm at non-trivial rates, highlighting the importance of explicit measurement of clinical safety.

20.
arXiv (CS.CL) 2026-06-11

Models That Know How Evaluations Are Designed Score Safer

The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings. Prior work has identified test-time contextual cues, such as hypothetical scenarios, as a source of verbalized evaluation awareness and subsequent behavioral shift. In this paper, we investigate a potential explanation of this phenomenon: evaluation meta-knowledge, defined as parametric knowledge about the structural traits that characterize evaluations. Similar to dataset contamination, where benchmark exposure leads to higher performance through memorization, we hypothesize that models trained on texts describing evaluation practices may implicitly learn to recognize and respond to evaluation-like contexts, for instance, through exposure to scientific articles or social media posts about AI benchmarking. To test this, we fine-tune models on synthetic documents describing evaluation traits such as verifiable structures or moral dilemmas. Evaluating this fine-tuned model on six safety benchmarks, we find that it is significantly safer than the base model and control model. This behavioral shift persists even when restricting the analysis to responses lacking explicit verbalization of evaluation awareness. Our results demonstrate that evaluation meta-knowledge may inflate safety benchmark performance, introducing a novel confounder that is independent of explicit memorization or verbalized evaluation awareness, thus, challenging to detect. These findings have important implications for the design and interpretation of AI safety evaluations. Our code and models are available at https://github.com/compass-group-tue/arxiv2026_evaluation_meta_knowledge.

21.
arXiv (CS.CL) 2026-06-11

ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing

Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded. We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor logits into the target model's vocabulary token-by-token at each decoding step; a small LLM judge then selects the safest among K candidate continuations. No weights are changed, and the safety-utility trade-off can be tuned at deployment without retraining. Across both cross-vocabulary and same-vocabulary evaluation pairs, ALIGNBEAM substantially raises refusal on adversarial benchmarks while keeping task accuracy and inference overhead within practical bounds. The results show that safety alignment can be transferred between model families at inference time, without touching either model's weights.

22.
arXiv (math.PR) 2026-06-18

A scaling limit theorem for controlled branching processes with a size-divisible term

arXiv:2508.17116v2 Announce Type: replace Abstract: This paper establishes general sufficient conditions for a sequence of controlled branching processes to converge weakly on the Skorokhod space. We focus on a class of control mechanisms that extend previous results by decomposing those random variables into the sum of two independent components: an immigration term, which depends on the current population size, and a size-divisible term, which can be expressed as the sum of random contributions from each individual. This extension allows us to capture a broad range of control functions including Poisson, binomial, and negative binomial distributions, commonly used in the literature. The assumptions are formulated in terms of probability generating functions of the offspring and control laws, distinguishing in this latter between the immigration and the size-divisible parts. The limit process is shown to be a continuous-state branching process with dependent immigration. The proof essentially relies on tightness arguments and the identification of a martingale problem. We also identify the special case in which the limit reduces to a classical Feller branching diffusion with immigration.

23.
arXiv (CS.CL) 2026-06-16

HK-LegiCoST: Leveraging Non-Verbatim Transcripts for Speech Translation

We introduce HK-LegiCoST, a new three-way parallel corpus of Cantonese-English translations, containing 600+ hours of Cantonese audio, its standard traditional Chinese transcript, and English translation, segmented and aligned at the sentence level. We describe the notable challenges in corpus preparation: segmentation, alignment of long audio recordings, and sentence-level alignment with non-verbatim transcripts. Such transcripts make the corpus suitable for speech translation research when there are significant differences between the spoken and written forms of the source language. Due to its large size, we are able to demonstrate competitive speech translation baselines on HK-LegiCoST and extend them to promising cross-corpus results on the FLEURS Cantonese subset. These results deliver insights into speech recognition and translation research in languages for which non-verbatim or ``noisy'' transcription is common due to various factors, including vernacular and dialectal speech.

24.
arXiv (CS.CL) 2026-06-12

TAB-PO: Preference Optimization with a Token-Level Adaptive Barrier for Token-Critical Structured Generation

Direct Preference Optimization (DPO) is an effective and widely adopted approach for offline alignment but is poorly matched to ontology-driven structured prediction, where preferred and rejected JSON objects often differ in only a few schema-defining tokens. In this low-edit-distance regime, sequence-level DPO spreads gradient mass across non-critical serialization tokens (gradient dilution) and can reduce likelihood on rare, under-confident preferred schema tokens (token erosion). To address these limitations, we first develop a confusion-aware preference-construction strategy that augments expert-curated ambiguity patterns with empirical structured-error modes estimated from validation-set SFT predictions, synthesizing minimally perturbed, schema-valid negatives that focus preference learning on realistic ontology-level decision errors. We then introduce Token-Adaptive Barrier Preference Optimization (TAB-PO), a post-SFT objective for token-critical structured generation. TAB-PO adds a confidence-gated token-level barrier that applies supervised anchoring to under-confident schema tokens. On the public SciERC scientific information extraction task, evaluated with Llama/Qwen models from 1.5B to 70B, TAB-PO improves ontology-critical semantic-label and relational-linking metrics over SFT by 11.59% on average, wins 100% of comparisons against the strongest token-level and sequence-level DPO variants on these metrics, and surpasses leading frontier models by 14.71%, while delivering strong gains in textual grounding.

25.
arXiv (CS.LG) 2026-06-16

When Does q-error Predict Plan Regret? Three Regimes of Cardinality-Estimation Error

arXiv:2606.15600v1 Announce Type: cross Abstract: Cardinality-estimation (CE) research ranks estimators by q-error, yet it is well known that q-error is an imperfect proxy for query-plan quality. We give a measurement-driven account of when it is a good proxy and when it is not, and why. Modeling plan selection as an argmin over a piecewise-linear cost landscape, we find that plan regret (the cost of the chosen plan relative to the optimal, under true cardinalities) is governed by plan-cost geometry in a regime-dependent way. (i) For small errors, a true-point condition number kappa predicts regret and out-predicts q-error; its predictive power decays to zero as error grows, as a local linearization must. (ii) For large errors – where deployed learned estimators operate – an estimator-independent average-case sub-optimality measure ACS-infinity predicts which queries are regret-prone (Spearman rho ~ 0.54 on STATS-CEB), while q-error is nearly uninformative at the query level (rho ~ 0.05). (iii) The worst case is Haritsa's maximum sub-optimality (MSO). The three are one cost-ratio spectrum under three weightings. We prove a limit law ACS-infinity = sum_k r_k pi_k with cardinality-independent combinatorial weights, and validate every claim on STATS-CEB and JOB-light with four released estimators under pre-registered decision rules, and confirm on real PostgreSQL runtime that ACS-infinity predicts regret where q-error does not. The contribution is conceptual and empirical – an average-case companion to worst-case robust query optimization, and a characterization of when an accuracy metric tracks plan quality – rather than a new estimator. Code and the full pre-registration are public.