In briefContinuous-Action World Models for Transferring Robot Dynamics Across Embodiments
SummaryThis paper introduces Continuous-Action World Models (CoAWM), a framework that learns spatiotemporal dynamics priors usable across different robot embodiments and control frequencies. It uses a shared latent function space, first pretrained on trajectories from heterogeneous robots and sampling rates, then connected to a world-model backbone that conditions on observations, language, and robot state. A single decoder returns end-effector actions at any queried time.
In briefScaling math proof training data using structured textbooks
SummaryThis paper introduces TOME-Math, a corpus converting 2,118 university mathematics textbooks into about 380K structured proof-supervision records that keep chapter order, proposition-proof links, dependency references, and source provenance. The authors use these structural signals to explore curriculum learning across mathematical subjects and to study scaling with supervision volume and source breadth. At full scale, the paper reports that fine-tuning on TOME-Math raises the unweighted mean score by 16.7 percentage points over the base model and exceeds three public instruction-data baselines in aggregate.
In briefCombining and guiding several diffusion models with a generalized Feynman-Kac correction method
SummaryThis paper introduces generalized Feynman-Kac correctors (GFKC), a framework for sampling from an arbitrary target function defined over several pretrained diffusion models. The authors show that geometric interpolation between an initial function and the target satisfies a Feynman-Kac PDE and derive its potential in closed form. The method runs in the log domain, supports reward-based steering, and corrects generation on the fly without retraining. It is validated on analytic Gaussian models and on diffusion models trained on MNIST.
In briefTesting AI language models on randomly chosen unsolved math conjectures
SummaryThe authors introduce OpenConjecture, a dataset of conjectures from arXiv mathematics papers published since February 2026. Frontier GPT models were applied to 200 randomly sampled conjectures over three passes, producing candidate solutions for roughly 60% of them, with counterexamples about 1.3 times more common than proofs. The paper also examines solution novelty, literature citation, and whether resolutions rely on errors or underspecification in the conjecture.
In briefLLM agent reasoning for predicting pediatric diabetes outcomes from multimodal medical records
SummaryThis paper tests large language model agents for clinical prediction in children, using records from 2,489 youths with type 2 diabetes and 4,018 with type 1 diabetes. Data included structured records, social determinants from clinical notes, and glucose monitoring data for type 1 patients. Using 18 months of follow-up, the authors predicted seven type 2 diabetes outcomes at 24 months. The evidence-citing CEDAR pipeline outperformed trained machine learning models in both cohorts, and two pediatric endocrinologists reached the same prediction as CEDAR in about nine of ten cases.
In briefHiFOR: Hierarchical Foresight for Robot Vision-Language-Action Models
SummaryThis paper proposes HiFOR, a framework for vision-language-action robot models that predicts future states at multiple time scales. A Hierarchical Dependency Mask jointly predicts short-term state changes, semantic milestones, and final goal states. The method organizes future latent representations along time, view, and spatial dimensions, and retrieves relevant future information guided by evolving action representations during denoising. Experiments were conducted on two simulation benchmarks and six real-world robotic manipulation tasks.
In briefThe Athlete as a Scientist of One: Agentic Bayesian Experimental Design for Individualized Physiology from Consumer Wearables
SummaryThis paper presents EpiCoach, an agentic system that studies one cross-country mountain biker as an auditable N-of-1 experiment. Claims about the athlete are kept in an evidence-gated registry, each requiring a probabilistic prediction before a session and updated only through logged statistical fits. Predictions come from a digital twin built from heart rate, GPS and barometric altitude, and the agent designs each next training week as a constrained experiment. Evaluated on eighteen months of data via replay benchmarks and a live prospective phase, the authors release the registry, benchmark and protocol.
In briefCurriculum or Easy Start? A Benchmark of Unsupervised Environment Design for Quadrupeds
SummaryThis paper proposes a benchmark that compares 14 unsupervised environment design, domain randomisation and robotics curriculum methods under one protocol on five 3D quadrupeds in MuJoCo MJX. Terrains are described by eight interpretable parameters, and all curricula train the same PPO student with matched environment-step budgets. No method ranks in the top group on every robot and level set, and methods with easier training levels often lead on in-distribution random terrains. On three robots the leader on random levels differs from the leader on held-out terrain.
In briefYour LR Schedule May Not Be Principled: Effective Learning Rate Scheduling in Weight-Normalized Training
SummaryThis paper argues that the pretraining loss gains of weight-normalized training mainly come from its effective learning rate trajectory rather than from normalization alone. It proposes AIR (Affine Inverse-square-Root decay), a schedule combining a 1/√t decay with an affine adjustment to meet set peak and terminal learning rates. In 128B-token Mixture-of-Experts pretraining, AIR reduces terminal training loss by about 0.01 relative to cosine, and applying the same effective learning rate trajectory to weight decay reproduces similar training dynamics.
In briefPathoJev: A pathology model that answers multiple slide questions efficiently from shared visual context.
SummaryPathoJev is a pathology model that directly predicts answer probabilities for several questions about the same whole slide or region, instead of generating separate text responses. It encodes each slide once and evaluates question suffixes in batches over the shared visual context, reusing recurrent states to limit added computation. Across three whole-slide benchmarks, its 9B and 27B versions reach mean accuracies of 77.2% and 78.1%, exceeding SlideChat by 5.4 and 6.3 points, and the 9B model runs 9.1 times faster than SlideChat on WSI-Bench.
In briefDon't Lose the Thread: Making LLM Agents Manage Their Own Memory on Long Tasks
SummaryThis paper addresses LLM agents' struggles on long-horizon tasks, where cluttered working memory dilutes attention and hinders planning. It proposes CORAL, a paradigm in which agents checkpoint task progress and verified facts, then reset cluttered context and resume from a clean state, trained with a multi-episode reinforcement learning algorithm. On the GAIA benchmark, an 8B model with CORAL reaches 42.8% average accuracy, using 6-7 times fewer input tokens than a ReAct agent.
In briefA unified Hessian-guided method for post-training quantization of language models
SummaryThis paper proposes a unified framework for post-training quantization of large language models. Using a tractable Fisher approximation, it derives a Hessian-guided estimate of the output change caused by joint weight and activation quantization. The method jointly optimizes quantization ranges, linear transformations, and weight rounding under a shared output-sensitivity objective. The authors report that it outperforms state-of-the-art transformation and rounding methods in both channel-wise and block-wise settings.
Mads Toftrup, Marco Federici, Boris van Breugel et al.
In briefCombinatorial Residual Network: a classifier that uses the decay chain to handle jet assignment without explicit labels
SummaryParticles that decay into several jets require each jet to be matched to its parent particle. The Combinatorial Residual Network (CoReNet) builds its representation along the assumed decay chain, evaluates every jet assignment with shared weights, and combines these hypotheses with learned weights, training on the classification loss alone. On simulated Higgs boson pair production in the bb̄bb̄ final state against QCD multijet background, its 1508 parameters outperform kinematic baselines. It also performs comparably to a transformer network over two orders of magnitude larger on HH→bb̄W⁺W⁻ against semileptonic tt̄.
John Alison, Patrick Bryant, Alejandro Gomez Espinosa et al.
In briefWizard: Real-time, agent-controlled manipulation of any virtual world
SummaryThis paper presents Wizard, an agentic game world model that enables real-time, precise control of character motions and generalizes zero-shot to new scenes, characters, and actions. The authors build a dataset covering 750 scenes and 178 characters, using skeleton and geometry information as structured representations to condition a video generation model. An adaptive masking mechanism improves alignment between these conditioning signals and generated video, while an agentic framework explicitly maintains and updates game states so players can freely configure their virtual worlds.
In briefRFW-A: Automatically generating 400K articulated 3D objects across 500+ categories using AI agents
SummaryThis paper presents RFW-A, a method where AI agents build reusable category-level generators that then produce articulated objects through procedural execution without further model calls. The authors use it to create RFW-A-400K, a dataset of over 400K simulation-ready articulated assets spanning more than 500 categories. Experiments report high structural and kinematic validity, and training on this data improves part-level articulation estimation.
WhyThe authors combine agentic program synthesis with procedural generation, reporting 400K articulated assets over 500 categories and zero-shot policy transfer to real objects.
In briefExplaining multistep, multivariate wind power forecasts with GradientSHAP
SummaryThis paper adapts the GradientSHAP method to explain multivariate, multistep wind power forecasting models, an area that has been largely unexplored. The authors create new visualizations that give both local and global explanations, showing relationships among environmental factors, site conditions, and temporal patterns. They find that with careful handling of SHAP values and domain expertise, the method yields useful insights into model behaviour. The authors also note that its complexity makes practical deployment difficult.
In briefHow pretrained language models jointly encode two independently changing temporal variables
SummaryThe paper studies how language models keep track of several contextual variables that change independently within one document. Using controlled documents with separate language-switch and topic-change timers, the authors find that both timers can be linearly decoded from the same hidden states across ten pretrained checkpoints. The results support an approximate factorization of these representations in overlapping subspaces.
In briefRadiology guideline-based decision graph for clinical decision support
SummaryThis paper introduces GDGraph, which uses a large language model to split radiology guidelines into sections and build an executable decision graph from extracted subgraphs. At inference, a fast model answers localized clinical questions while deterministic traversal resolves Boolean logic, producing a traceable decision path. On 347 cases covering renal cyst, adnexal mass, and thyroid nodule guidelines, accuracy rose from 57.3% to 90.8%.
In briefHow to Loop a Transformer: Scaling, Specialization, and Parallel Inference
SummaryThis paper studies looping in transformer models, which reuses the same weights by applying a transformer, or part of it, repeatedly to add depth without adding parameters. The authors train dense decoder-only models up to 1B parameters, varying which parts are looped, the number of loops, and how information passes between loops. They find that looping the full model is strongest at a fixed parameter budget, while looping a middle block is better under limited training compute, and that gains from more loops can saturate or become unstable. They also add loop-specific low-rank adapters and modify Parallel Loop Transformers so that quality is preserved while decoding runs in parallel.
Chanakya Ekbote, Artidoro Pagnoni, Devendra Singh Sachan et al.
In briefMage-VL-Cowork: A Full Training Recipe for Hybrid Computer-Use Agents Learned from Evolved Harnesses
SummaryThis paper presents Mage-VL-Cowork, a 4B-parameter computer-use agent that combines visual GUI interaction with code and terminal execution in one policy. It is trained through a five-stage curriculum spanning midtraining, supervised fine-tuning, and reinforcement learning. The authors also introduce Mage-CUA-Forge, a workflow that synthesizes 293,806 desktop and web tasks across 6,915 skills, and evolve the teacher harness to produce high-quality trajectories. The model reaches 66.0% on OSWorld-Verified, which the authors report as state-of-the-art at its parameter size.
In briefAccuracy and calibration of Decision and System One AI models on rheumatology exam and diagnosis questions
SummaryThis study evaluates how accurately several Decision and System One models answer rheumatology examination questions and diagnostic case vignettes, and how well their confidence matches their correctness. On Spanish specialty-training exam questions, most models scored highly, with Jev at 93.0% in Spanish and 91.9% in English, below GPT-6 Sol at 99.4%. On diagnostic vignettes, all System One models scored below GPT-6 Sol at 76.9%, and Jev reached 50.0% while showing overconfidence.
Alfredo Madrid-Garc\'ia, Diego Benavent, Fabian Lechner et al.
In briefVampire-Hammer: calling an automated prover from the Lean proof assistant
SummaryThe paper presents Vampire-Hammer, a tool that lets the Lean interactive theorem prover hand goals to the Vampire automated theorem prover and then checks the resulting proofs with the Lean kernel. It processes goals and premises with the Lean-auto framework and uses standard TPTP and annotated TSTP formats to support proof reconstruction. On 196,083 eligible Mathlib theorems with a 10-second budget, it solves 21,158 problems that the other evaluated tools did not solve within the same time.
In briefTextual Response-Aware Guidance for Zero-Shot Adversarial Robustness in Vision-Language Models
SummaryThis paper studies how to make vision-language models more resistant to adversarial attacks while keeping their zero-shot performance. It argues that two common guidance methods each have limitations, and proposes TRAG, which constrains visual representation drift using the pre-trained text encoder's responses to local input changes. Experiments across architectures and tasks report state-of-the-art clean and robust performance, and further analysis indicates TRAG helps preserve pre-trained image-text relationships.
In briefRimuru: Unified Memory Management for LLM Serving via Elastic Expert Offloading
SummaryThis paper addresses device memory pressure in serving large language models, where fixed memory partitions for model weights and attention states cannot adapt to changing workloads. It presents Rimuru, which uses a shared page pool, memory views with per-tensor page sizes, a multi-level page allocator, and a page-table-aware fused MoE kernel to share memory elastically across heterogeneous tensors. A workload-aware adjuster reallocates pages at runtime. Implemented on vLLM, the system is reported to achieve up to 19.03x baseline throughput and a 98.1% average reduction in TTFT.
In briefDeep learning with causal constraints for early warning of rapid cyclone intensification in the western North Pacific
SummaryThis paper presents CausalVisionRI, a deep-learning system that warns of rapid intensification of tropical cyclones over 24 hours in the western North Pacific. It combines infrared satellite imagery, atmospheric and oceanic fields, and a causal graph discovered from observations as a training prior. On a forward holdout test covering 2020 to 2024, its five-member ensemble achieves a Brier skill score of 0.259, outperforming several reference systems. Ablation experiments show that the causal graph pathway improves performance.
In briefSelf-training with contrast-frame subtraction for coronary X-ray vessel segmentation across hospitals
SummaryThis paper addresses the drop in performance of coronary X-ray angiography vessel segmentation models on data from centres not seen during training. The authors generate pseudo-labels from nnU-Net predictions and a contrast map obtained by subtracting a pre-contrast frame, then self-train without extra manual annotation. Self-training improved results on external datasets, and the size of the gain depended largely on the sensitivity of the model that produced the initial pseudo-labels.
In briefMAServeBench: Benchmark for Serving Multi-Agent LLM Systems with Hierarchical Collaboration Workloads
SummaryThis paper presents MAServeBench, a benchmark for evaluating serving systems that run multi-agent large language model applications. It treats how agents collaborate and exchange context as a configurable part of the workload, so the same tasks can produce different serving demands. The benchmark records realized operations and dependencies for reproducible comparisons and links them to hardware behavior, showing why higher serving efficiency does not always mean tasks finish sooner.
Shengxuan Qiu, Haochen Huang, Shuzhang Zhong et al.
In briefDolphinRouting: Cluster-Level Routing for Split LLM Serving
SummaryThis paper presents DolphinRouting, a method for routing requests in large language model serving systems that separate prefill and decode phases. Requests are grouped into clusters by prompt length, cache-adjusted input lengths, and output-length information, and each cluster learns a binary choice between local and remote prefill from end-to-end SLO feedback. Evaluated on a real disaggregated cluster with public benchmarks, a mixed workload, and real data, the k-means variant reduces P90 TTFT by 5.7 to 7.9% and P90 TPOT by 2.6 to 10.1% compared with Dynamo.
In briefOnline knapsack with removable items using recourse and size bounds
SummaryThis paper studies the online unweighted knapsack problem, where items arrive one by one and must be accepted or rejected immediately to maximize the total size of accepted items. It allows previously accepted items to be removed and considers two extensions: knowing an upper bound on item sizes and being able to reinsert previously rejected items. The authors establish close upper and lower bounds on the competitive ratio under these combined assumptions. They also introduce adaptive online algorithms that do not require knowing the size upper bound yet achieve comparable performance guarantees.
Dimitris Fotakis, Laurent Gourv\`es, Aris Pagourtzis et al.
In briefOnly Accepted Tokens Count: Energy Accounting of Speculative Decoding Under Latency Targets
SummaryThis paper examines the energy cost of speculative decoding, in which a drafter proposes several future tokens that a target model verifies in one forward pass. The energy saving depends on how many drafted tokens are accepted, which varies by model, drafting scheme, and text. Measured on H100 and H200 GPUs with SGLang at batch sizes up to 64 across 5 corpora, the saving shrinks as batch size grows and can become a loss at large batches. Speculation only adds energy savings when acceptance is high.
In briefLoRe-MIM: A self-supervised method for learning features from whole-body PET/CT scans using masked image reconstruction
SummaryThis paper addresses the shortage of expert annotations for whole-body PET/CT deep learning by assembling a multi-tracer, multi-institutional dataset of 6,531 scans. It proposes LoRe-MIM, a self-supervised framework that learns multimodal representations by reconstructing masked images, where masked regions are either zeroed out (Lost) or replaced with the corresponding complementary imaging modality (Replaced). The learned representations are evaluated on tasks including multimodal reconstruction, cancer versus non-cancer patch classification, whole-body lesion segmentation, label-efficient training, and robustness assessment. The authors report that LoRe-MIM consistently outperforms corresponding state-of-the-art baselines across these tasks.
Amartya Bhattacharya, Bashirul Azam Biswas, Biratal Raj Wagle et al.
In briefGuiding latent reasoning exploration with potential-based trajectory balance
SummaryThis paper introduces SubLAT, a method for continuous latent reasoning in language models that combines a relative answer potential with Subtrajectory Balance to learn stochastic latent exploration. The model is trained jointly with answer supervision, CoT supervision, and representation distillation. Experiments on arithmetic and deductive reasoning show that solution coverage increases with larger sampling budgets and improves on several structural out-of-distribution benchmarks.
In briefSelf-supervised brain encoders that align across people in fMRI
SummaryThe paper adapts a joint embedding predictive architecture (JEPA) to learn participant-specific fMRI encoders from brain data alone, without ever seeing stimulus images. On the Natural Scenes Dataset (NSD), the embeddings achieve competitive results on brain captioning and image reconstruction against supervised pipelines that use images and language. The authors also show that encoders for different participants can be aligned fully unsupervised, so a decoding head trained on one participant transfers to another without fine-tuning while keeping most of its within-subject performance.
Serwan Jassim, Victoria Bosch, Daniel Anthes et al.
In briefNeighborhood Elites: Quality-Diversity Search for Detecting Wrong Answers in QA
SummaryThis paper addresses the difficulty of detecting wrong answers from large language models in question answering. It proposes neighborhood elites (NE), a method that uses quality-diversity search to build answer-preserving input rewrites spread across two axes: the kind of edit and how explicitly the context supports the answer. Across five extractive-QA datasets and five instruction-tuned LLMs, NE improves failure detection over other neighborhood constructors in at least 24 of 25 pairs, and also outperforms semantic entropy computed over decoder samples.
Mesut Ceylan, Arda Baris Basaran, Elgar Fleisch et al.
In briefTaskFrame: A Framework Linking Task Planning to Frame-Level Control for Game Agents
SummaryThis paper presents TaskFrame, a hierarchical framework for game agents that connects task-level planning with frame-level keyboard and mouse control. A vision-language planner breaks a high-level goal into ordered atomic instructions, while a causal Transformer controller predicts actions at each game step using the active goal and recent observations. Experiments in Minecraft show competitive instruction decomposition compared with general-purpose vision-language models, and a higher combat success rate than the evaluated baselines in closed-loop evaluation.
In briefFrom MS/MS Evidence to Adaptive Spectroscopy: Reinforcement Learning Guided by Candidate Structures
SummaryThis paper proposes a structure-informed deep Q-network (SI-DQN) that uses existing MS/MS evidence to decide whether further characterization is needed, which measurement to request, and when to stop. The policy was evaluated on a Multimodal Spectroscopic Dataset containing simulated MS/MS, infrared, and NMR spectra. Results show higher Top-1 accuracy with far fewer additional measurements than complete acquisition, and lower composite loss than a rule-based policy.
In briefLearned terminal costs enable short-horizon MPC for autonomous trucks in busy highway traffic
SummaryThis paper addresses autonomous highway driving for heavy-duty trucks, combining Model Predictive Control (MPC) with Reinforcement Learning (RL). A finite-horizon MPC solved by dynamic programming uses a pre-trained RL model to estimate terminal cost. The method is tested in simulation with reactive surrounding traffic against pure MPC with zero terminal cost, rollout with a base-policy terminal cost, and pure RL. The hybrid approach achieves zero collisions and the lowest cost among all configurations.
Hooman Hamze, Benjamin Houshmand, Deepthi Pathare et al.
In briefWhen Requirements Interact: Long-Horizon Coding via Interdependent Specs
SummaryThis paper introduces CodeSpecWeave, a framework that synthesizes coding tasks with rich, interdependent specifications grounded in real developer seeds. A synthesis agent creates coupled requirements and executable tests, while a review agent audits solution trajectories from multiple models to ensure test reliability. The authors built 4,670 training tasks across 2,865 repositories and used reinforcement learning on them, substantially outperforming the base model and existing baselines on DeepSWE-v1.1. They also released CodeSpecWeave-Bench, a benchmark of 100 tasks and 1,044 specifications.
In briefStatistical Analysis of Verified Synthetic Data in Linear Regression
SummaryThis paper studies self-training on synthetic data within a linear regression framework. A pre-trained estimator generates data with injected noise, a verifier filters samples by a threshold, and the model is then fine-tuned on accepted samples. The authors derive finite-sample risk bounds across four regimes to show how verification threshold, generation noise, and pre-training error affect prediction risk.
In briefEstimating ground speed from Doppler spectra with per-frame confidence
SummaryThis paper proposes a method for estimating speed over ground from ground-facing Doppler spectra, aimed at mobile platforms such as railway systems. It builds on an autoencoder reconstruction approach, adding a speed estimation head on the latent features and a lightweight confidence head used at inference, with noise augmentation training. Experiments show improved speed estimation accuracy over classical baselines and confidence estimates that indicate the reliability of individual predictions under adverse signal conditions.
In briefConformal annotation with randomized expert review for online cell-type labeling
SummaryThis paper proposes CARED, a distribution-free framework for AI-assisted cell-type annotation under distribution shift. It randomly selects cases for expert review and uses the resulting labels to calibrate later decisions. The authors prove long-run control of cumulative error without stationarity or exchangeability assumptions. Experiments on pancreas, lung, and ovary annotation streams show that up to 70% of cells can bypass expert review while errors remain controlled.
In briefDistribution-calibrated, budget-neutral token credit for multimodal RL with verifiable rewards
SummaryThe paper proposes TACE, a two-stage estimator for advantage computation in GRPO. The first stage maps rewards to a standard normal via a one-dimensional optimal-transport map to align reward distributions across tasks and cap outliers. The second stage rank-calibrates token-level evidence weights within each trajectory and renormalizes them to mean one, conserving total credit and sign. On a multimodal reasoning suite, TACE achieves the best results on most benchmarks but not in every setting.
In briefReaction precedents let retrosynthesis prediction go beyond its training data
SummaryThis paper examines how far retrieved reaction precedents can make up for gaps in a retrosynthesis model's task-specific training coverage. Using 2.71 million reactions, the authors limited training and retrieval by transformation type or time, then expanded only the retrieval corpus while keeping model parameters fixed. Precedents substantially raised accuracy on transformations excluded from training and improved all 14 general-purpose models, with the best surpassing eight specialized models.
In briefLLM-guided planar loco-manipulation for quadruped robots using language planning and pushing policies
SummaryThe paper addresses robotic navigation in cluttered environments where reaching a goal may require physically moving obstacles. A language-conditioned task planner, fine-tuned with LoRA on synthetic data, generates variable-length sequences of parameterized actions executed by low-level reinforcement-learning policies. The planner outperforms a classical planner and a prompted frontier model on interactive navigation tasks. An end-to-end whole-body planar pushing policy reaches a 93.3% success rate in simulation, compared with 83.5% for a hierarchical baseline.
Georges Jetti, Michael Khayyat, Jean-Pierre Sleiman et al.
In briefEnvOS: Auditing AI computer agents in software that changes mid-task
SummaryThis paper introduces EnvOS, a framework for building and certifying environments whose software changes while an AI agent is working on a task. It checks before a run that changes fire correctly and tasks remain solvable, and re-checks verdicts from archived evidence afterward. Tested on stale-information scenarios in clinic booking and insurance-claim approval, the results show that whether an agent notices a change depends on the model and the interface shown to it.
In briefJudged in Company: Batching Items in One Prompt Makes LLM Verdicts Depend on Neighbours
SummaryThis paper examines whether asking a language model to grade many independent items in one prompt, a common cost-saving practice, changes the verdicts. Across three judgement tasks and five open-weight models, placing the same 2,000 items into different random batches changed 11-22% of decisions on harm and grading. Most of this variation comes from the order and position of neighbouring items, and majority voting over permutations does not remove it. The authors recommend grading ambiguous items one per prompt and reporting agreement with single-item grading when batching.
WhyThe authors show batched grading makes verdicts depend on co-batched items, measuring flip rates across three tasks and nine models and decomposing them.
In briefScienceJudge: Pattern-Based Data Augmentation for Checking Sentence-Level Scientific Claims
SummaryThis paper addresses checking whether each sentence in an AI-generated scientific paragraph is supported by a given multimodal evidence bundle. The authors propose Pattern Perturbation Augmentation (PPA), which models reporting errors as faithfulness-breaking transformations to create hard negative examples. Their ScienceJudge system combines vision-language and LLM predictors, raising the local development-set average score from 0.8124 to 0.8692. It ranked first on the official NLPCC 2026 Task 10 Track 1 leaderboard with a final score of 0.6890.
In briefMethod for certified behavioral retention during continual model adaptation using held-out anchor data
SummaryThe paper introduces CABA, a post-update gate that keeps historical observations aside from candidate training and compares old and new policies on frozen representations. It accepts the largest tested interpolation coefficient whose worst-anchor forward KL divergence stays within a declared budget. In a UCI Human Activity Recognition study, at budget 0.05, mean forgetting fell from 1.01% to 0.00% while final average accuracy dropped from 93.04% to 88.48%.
In briefMining translation pairs from linguistic archives for low-resource languages
SummaryThis paper proposes extracting parallel translation examples from linguistic archival materials such as grammars and dictionaries, which cover languages with little online presence. The authors build a pipeline combining an OCR model and a vision language model so that extracted text stays aligned with visual layout information. The extracted bitexts are used across 20 languages to train machine translation models and to support in-context learning for large language models. For in-context learning, the reported chrF2++ gains over zero-shot translation are +15.78 for low-resource and +6.94 for high-resource target languages.
Chutong Meng, Alec J. Bateman, Antonios Anastasopoulos
In briefEpiAgent: Epistemic Self-Play Framework for LLM Agents
SummaryThis paper proposes EpiAgent, a framework that gives LLM-based agents a learnable layer for maintaining beliefs about others' private information. The policy is split into an Epistemic Update stage and an Action Planning stage, optimized alternately through a bi-level formulation called Epistemic Self-Play, using an intrinsic prediction reward. The authors report that EpiAgent outperforms prompting-based and RL-based baselines in mixed and competitive scenarios.
