Bridge-LP is a new framework for link prediction on text‑attributed graphs that combines Graph Neural Networks and Language Models. It introduces three design components: jumping connections for robust language model training, compressed embeddings to encode neighborhood information, and a cache embedding strategy to reduce computational cost. Experiments on nine datasets show that Bridge‑LP outperforms 40 existing models, achieving up to a 43.06% improvement in Mean Reciprocal Rank.
By Chen Shao, Ruirui Wang, Tobias Kafer, Danai Koutra
The paper introduces SkillCompiler, a framework that compiles self‑evolving agent skills into deterministic programs to reduce uncertainty from LLM reasoning. It decomposes skills into fine‑grained operations, selectively compiles well‑specified parts into validated programs, and reuses these programs across evolving skills. Experiments on five benchmarks and six model families show SkillCompiler outperforms state‑of‑the‑art methods, achieving up to a 32.33% improvement in task performance.
By Chengyu Yang, Yunxiang Zheng, Zongzheng Li, Jiamei Li, Boxun Lan, Xiaoyu Liu, Xiaoyu Zhang, Xiaohong Chen, Ming Hu
EvoTeam is a collaboration‑centric evolution framework for large‑language‑model based multi‑agent systems. It jointly evolves team organization, interaction rules, and individual agent skills using execution experience, while maintaining a library of reusable teams and guiding collaboration through team‑specific interaction rules. Experiments on four benchmarks show its effectiveness, and the code is publicly available on GitHub.
By Qinuo Cao, Xin He, Dianxuan Fu, Lizhi Zhang, Zhaoqing Xing, Jiapu Wang, Xin Wang, Na Zou, Yi Chang, Qinggang Zhang
AegisFlow is a training‑free framework that guides generative models, such as diffusion and flow matching, along safe paths by formulating the problem as a global optimal control task. It incorporates a time‑varying barrier function into the cost functional, enabling non‑myopic enforcement of constraints throughout the entire generative trajectory. Experiments show that AegisFlow outperforms existing methods in text‑guided image manipulation and robotic planning, achieving a better balance between target alignment and constraint satisfaction.
By Kunpeng Liu, Boshi Zhang, Keyou You
The paper introduces SPLICE, a method that identifies and closes information gaps in multi‑agent workflows for web‑task completion. SPLICE locates missing evidence after an initial run, adds a single directed edge to deliver the evidence, re‑executes affected agents, and uses a selector trained on process features to decide whether to adopt the new answer. Experiments on five benchmarks show SPLICE improves web‑tool use scores by 8.0% and reasoning/planning scores by 5.7% with modest cost increases and zero‑shot transfer to new backbones.
By Yaxin Gong, Chongming Gao, Gangyi Zhang, Chenxiao Fan, Leyang Shen, Jiakai Wang, Dong Wang, Yang Liu, Haoyan Liu, Xinting Hu
The paper introduces a benchmark of 250 expert-curated questions on inertial confinement fusion to evaluate language models on scientific fluency, physical reasoning, and quantitative analysis. It compares various closed‑source, open‑source, and a specialized model called Prospero, showing that Prospero improves by 10.7 percentage points and that frontier models still make critical physics errors. Retrieval assistance helps smaller models but not frontier ones, and two stress tests highlight the need for expert assessment of physical assumptions.
By P. B. Radha, D. D. Meyerhofer, M. Bhattarai, B. Cheng, L. Diaz, C. L. Rousculp, A. G. Seaton, G. Gupta, P. Gurumurthy, V. Gutta, A. Jhunjhunwala, A. Vem, P. Cook, G. Rowe
The paper introduces Rewindable Interactive Proofs (RIPs), a new model for scalable oversight where a weak verifier can rewind a powerful AI prover to an earlier state during interaction. RIPs enable verification of oracle‑aided computations that are robust to small changes in oracle answers, achieving near‑linear verifier time without relying on cryptography. The authors provide a construction, implement it on MNIST, and demonstrate that RIPs can outperform existing proof systems in practice.
By Ziyi Guan, Yael Tauman Kalai, Ron Rothblum, Zoe Xi
The paper investigates the limits of predicting how many tokens a large language model will generate using only prefill-phase hidden states. It shows that current benchmarks mix predictor quality with task difficulty, leading to unstable accuracy when models, datasets, or sampling settings change. By modeling decoding as a stopped stochastic process, the authors derive bounds and an affine variance law that explain the observed normality of per‑prompt decode lengths and confirm that system error dominates model error.
By Haowen Shi, Xuzhe Yan, Fahao Chen, Peng Li
The paper introduces the Native Predicate Neuron, a differentiable logic‑reasoning module that can be inserted into standard neural architectures. It learns to alternate between logical operations using a single trainable parameter and employs a dual‑nature gradient mechanism that performs hard logical operations during training while remaining differentiable for backpropagation. Experiments on Fully‑Connected, Convolutional, Recurrent, and Transformer models show that adding Predicate Neurons speeds convergence, improves accuracy, and reduces loss, especially on tasks involving monotonic reasoning or discontinuities.
By Ioannis Konstantoulas, Iliana Loi, Kyriakos Sgarbas
The paper introduces a six‑module scanner for detecting unintended local and composite vulnerabilities in LLM‑based agent skills. Local Analysis achieved high recall and precision on benchmark datasets, outperforming existing scanners, while Composite Analysis accurately modeled inter‑skill dependencies and identified composite flaws. The authors validated findings with proof‑of‑concept attacks, disclosed cases to maintainers, and demonstrated effective mitigations that preserved legitimate functionality.
By Parth A. Gandhi, Avital Aviv, Prasanna N. Wudali, Yuval Elovici, Asaf Shabtai
The Prestige is a video‑understanding benchmark that tests cognitive visual reasoning using professional magic performances, where misdirection forces models to track events and explain surprising outcomes. The benchmark includes open‑ended questions covering temporal, causal, and robustness aspects, and reveals a large performance gap between state‑of‑the‑art VLMs and humans. Systematic differences between human and automated grading, recurring model failure modes, and insights from ablations suggest that current VLMs can answer multiple‑choice questions correctly yet still fail to generate accurate explanations, highlighting the need for better models and stricter evaluation methods.
By Shuo Wen, Beixi Du, Edwin Meriaux, Junming Shi, Chloe Si, David Meger, Doina Precup, Gregory Dudek
The paper introduces GUIDER, a method that jointly trains a decision flow controller and a reasoning policy to improve clinical guideline reasoning. GUIDER generates multiple plausible decision flows from guideline supervision, computes process rewards across these flows, and gates them on answer correctness to address both path–answer gaps and path multiplicity. Experiments on MedGUIDE and out‑of‑domain benchmarks show that GUIDER outperforms existing reinforcement learning baselines, achieving higher weighted accuracy and guideline adherence.
By Gaurav Bhatt, Marinka Zitnik
MAS-Loop introduces an execution‑grounded framework that couples an Orchestrator with a trajectory‑conditioned Refiner to close the feedback loop in training‑time multi‑agent system design. For each task, the Orchestrator builds and runs a source MAS, then the Refiner uses the resulting trajectory to locate failures, diagnose them, and propose bounded patches to agents or execution settings. The updated MAS is re‑executed, and the pairwise comparison of outcomes provides learning signals that enable both components to co‑evolve, yielding stronger systems and a new dataset, MAS‑JudgeCorpus, for trajectory‑aware MAS research.
By Hehai Lin, Qi Yang, Chengwei Qin
The paper introduces Negative Reinforce Dynamics (NRD), a new training approach for Generative Flow Networks (GFlowNets) that addresses the exploration–correctness trade‑off inherent in standard semi‑gradient updates. By linking the full gradient to a min‑min dynamic between a GFlowNet and a decoupled sampler, NRD effectively reduces balance errors in high‑error regions while preserving a single‑policy framework. Experiments on standard benchmarks show that NRD speeds up convergence and improves exploration efficiency compared to existing methods, without requiring a separate teacher or auxiliary policy.
By Pedro Dall'Antonia, Daniel Csillag, Diego Mesquita
Cord is a lightweight, prompt‑conditioned policy that augments confidence‑based decoding for masked diffusion language models (dLLMs). It learns geometric residuals—considering generation direction, proximity to revealed context, and two‑sided infilling—to decide when and how to deviate from the default confidence rule, sampling a single low‑dimensional coefficient matrix per rollout to generate a state‑dependent reveal trajectory. In experiments across three dLLMs and five benchmarks, Cord raises average task success from 32.0% to 48.2%, adds only 0.53 M parameters, and incurs a 5.39% inference latency overhead.
By Chengwei Zhou, Zhaoyan Jia, Massoud Pedram, Gourav Datta
The study evaluates six state‑of‑the‑art large language models on 400 adversarial prompts in Yoruba and Nigerian Pidgin, categorised into four harm types. Native‑speaker annotators scored responses and assessed reliability with Cohen’s κ, revealing that many models failed to understand the prompts, leading to low rejection rates in the African languages. Harmful content was frequently produced, with attack success rates higher in Nigerian languages and significant cross‑lingual safety gaps, especially for GPT‑4o Mini and Aya 101.
By Paula Chikaodinaka Anyanwu, Valdini Douglace Lemofouet, Blessing Ngozi Uzor, Ifeoluwa Dorcas Awotunde, Danielle Blanche Kapsa, Tassallah Abdullahi, Clemencia Siro, Idris Abdulmumin, Seid Muhie Yimam, Shamsuddeen Hassan Muhammad
The paper introduces BedBench, a benchmark of thirteen procedurally generated tasks designed to test vision‑language models’ ability to recover and consistently use exact 3D object properties. Across twelve open‑weight and four frontier models, the study finds a stark gap: humans achieve 93.3 % accuracy on per‑color counting, while the best model scores only 7.2 %. Even targeted fine‑tuning can raise a model’s accuracy to 68.6 % but leaves its answers about the most common colors inconsistent with its own per‑color counts, highlighting that recovering a property does not guarantee its consistent use.
By Khai Le-Duc, Nguyen Kim Hai Bui, An Ngo, Nam D. Hoang, Yueqi Song, Jiarui Liu, Thai Quoc Hoang, Fang Wu, Chris Ngo, Terry Jingchen Zhang, Mona T. Diab, Ran Xu, Silvio Savarese, Xiao-Ming Wu
The paper introduces EpiCoach, an agentic system that treats a single athlete as the subject of accountable N‑of‑1 science. It uses a mechanistic digital twin—comprising a lagged heart‑rate model, a fitness‑fatigue model, and a latent skill state—to generate probabilistic predictions for future training sessions, which are logged in an evidence‑gated hypothesis registry. Over eighteen months, the system’s predictions were evaluated both retrospectively and prospectively, showing improved heart‑rate forecasting and a transparent, auditable loop for personalized training guidance.
By Behnam Moradi
The paper introduces AnyGoal, a benchmark that tests world models on goal‑shifted task variants where state and action factors are recombined beyond the combinations seen in expert rollouts. It evaluates how well different world‑model families generalize to these unseen recombinations across multiple environments, revealing consistent failures. The study analyzes dynamics and representation shortcomings to identify where current models succeed and where they struggle in goal‑shifted prediction and planning.
By Heejeong Nam, Divyat Mahajan, Quentin Le Lidec, Randall Balestriero, David Klindt
PepEVOLVE is a dynamic, position‑aware framework for optimizing macrocyclic peptides that learns where to edit and how to improve multiple properties. It enhances pretraining with dynamic masking and CHUCKLES shifting, employs a multi‑armed‑bandit router to identify high‑reward residues, and uses an evolving optimization algorithm with group‑relative advantage to stabilize updates. In in‑silico tests on a Rev‑binding macrocycle benchmark, PepEVOLVE surpassed PepINVENT, achieving higher permeability‑weighted scores and converging faster while respecting structural constraints.
By Trieu Nguyen, Hao-Wei Pang, Shasha Feng