In briefTALLOP: Jointly tuning local training and communication compression for federated learning under time limits
SummaryThis paper addresses federated learning, where slow clients delay global synchronization, by proposing TALLOP, which jointly optimizes local training workload and communication compression under a per-round time constraint. It predicts achievable training loss under different computation budgets from early mini-batch loss trajectories (LENS) and estimates compression-induced degradation with a compression-error model (CLEN). Experiments in heterogeneous FL settings show TALLOP generally achieves competitive or improved time-to-accuracy compared with methods that optimize only local training or neither, though the advantage over training-only adaptation varies by setting.
In briefGOLD: Structure-aware RNA design method using latent diffusion with side-chain geometry
SummaryThis paper proposes GOLD, a method for designing RNA sequences that fold into specific 3D structures. It uses a GeoVAE to map discrete RNA sequences and continuous side-chain angles into a unified latent space, and applies gradient guidance from physical constraints during sampling. The authors report state-of-the-art results on sequence recovery, secondary structure validity, and 3D foldability, and state that code and data are publicly available.
In briefA100: A hybrid vision model for analyzing farm plant images
SummaryThis paper introduces A100, a vision foundation model for agriculture, pre-trained with self-supervised learning on 3 million unlabeled field images covering more than 100 crop species. It uses a hybrid CNN-Transformer backbone and, with a frozen backbone, outperforms ImageNet-12K supervised features on crop classification and leaf counting. Using only 1% of labeled data, it reaches 98.01% accuracy on plant classification, and the authors plan to publicly release the pretrained model.
Mayur Pratim Das, Ajai Dubey, Aksshay Mathew P et al.
In briefPersonaCUABench: Testing computer-use AI agents with users who phrase and answer requests differently
SummaryThis paper introduces PersonaCUABench, a benchmark of 74 tasks across 11 domains where only the requesting user changes, simulated by 44 personas. A persona agent restates each instruction in the user's own words and answers the agent's questions. Across nine backbones and 3,330 runs, mean task completion drops 36% relative to the original requests. The authors argue agents should be judged on whether they meet users' stated requirements and constraints, not just on task completion.
In briefFrom Approval to Abuse: Privilege Escalation Attacks on Harness-Protected Coding Agents
SummaryThis paper studies privilege escalation risks in coding agents protected by execution harnesses. The authors show that adversarial context planted in a repository can steer an agent during a routine task to request a plausible but overly broad allow-rule. Once a user approves it, later malicious actions matching the rule run without prompting. Experiments on platforms including Claude Code and Codex CLI report high success rates.
WhyReveals a novel risk where untrusted context steers allow-rule approvals in coding agents, quantified on production platforms with high success rates.
In briefPrism Attention: A proposal-refined index sharing method for faster LLM inference with long contexts
SummaryDecoding with long contexts makes large language models expensive, because each step must attend over a growing key-value cache. The paper notes that existing cross-layer sparse index sharing loses accuracy, since attention patterns change across layers. The authors propose Prism Attention, which replaces fixed index reuse with a coarse-to-fine strategy that keeps layer-wise adaptability while retaining sharing efficiency. The paper reports state-of-the-art results among mainstream sparse attention methods and a 3.4x end-to-end decoding speedup over FlashAttention at 128K context length.
In briefSelf-evolving AI agents can fix failures but struggle to generalize from them
SummaryThis paper studies whether large language model agents can improve themselves when they start without any accumulated experience and learn only from failed interactions. Failure-derived experience helps repair previously failed tasks but rarely produces sustained gains on unseen tasks. The authors find agents often misinterpret failures, and even verified-correct experience is hard to transfer. Successful trajectories substantially improve transfer, while verifier feedback and extra interaction give limited gains.
In briefUniField++: A unified AI framework that enhances MRI images across field strengths by repeatedly refining high-risk brain regions
SummaryThis paper presents UniField++, a two-agent framework that improves MRI image quality across multiple field-strength transitions, including 64mT-to-3T and 3T-to-7T. A VLM-based planner, trained on radiologists' annotations of hallucination-prone regions, predicts regional risk and forms restoration plans. A flow-matching enhancer follows these plans, repeatedly refining high-risk regions, and the authors report gains in PSNR, SSIM, and downstream whole-brain segmentation over comparison methods.
In briefWhen Does a Vision-Language Model Realize It Cannot See? Accuracy, Confidence and Abstention Versus Physical Visibility
SummaryThe paper measures five small open vision-language models on images degraded by fog and underwater conditions, using visibility in metres as the axis. Accuracy drops to chance within a narrow visibility band, while most models keep high confidence, creating a blind spot. The ImageNet-C fog severities correspond to visibility well above this failure band, so they do not reveal the problem.
In briefKiranaBench: Can a Language Model Run a Neighbourhood Shop?
SummaryThis paper introduces KiranaBench, a simulated Indian general store in which four roles (manager, cashier, helper, and bookkeeper) are each played by a separate LLM agent. Agents interact with customers and suppliers in English, Hindi, or Hinglish, with demand calibrated to India's national Household Consumption Expenditure Survey. Evaluating four open-weight LLMs, the authors find all fall far short of a fixed-rule shopkeeper baseline, with the best earning one-eighth of its profit and two finishing below an unmanaged shop. Performance is also lower for Hindi and Hinglish than for English supplier communication.
Bhuvan Arora, Ankita Karmakar, Chaitanya Jain et al.
In briefSink-Sparse Hybrid Attention: A Train-Free Way to Speed Up Text-to-Image and Text-to-Video Diffusion Transformers
SummaryThis paper addresses the high cost of dense self-attention when diffusion transformers generate images and videos from text. It proposes Sink-Sparse Hybrid Attention, a train-free method that reuses layer- and head-specific high-importance attention masks found from a small offline calibration set, combined with a lightweight pooled-KV branch. Model weights and sampling schedules stay unchanged. The authors report end-to-end speedups of 1.54x on Wan2.1-T2V-1.3B and 1.57x on Z-Image-Turbo while maintaining generation quality.
In briefDiffusionRVGE: Regression-Based Value Gradient Estimation for Reinforcement Learning of Diffusion Models
SummaryThis paper introduces DiffusionRVGE, a method that improves diffusion models with reinforcement learning by regressing centered rewards on centered final denoised images. It avoids reward derivatives and backpropagation through the sampling trajectory, lowering training cost. Trained jointly on five rewards, one SD3.5-Medium model matches or exceeds DiffusionNFT on seven of eight metrics at about one fifth of its estimated compute. In single-reward training, it reaches matched reward thresholds 2.1 to 6.7 times faster than a reproduced DiffusionNFT.
WhyRegressing centered rewards on clean images avoids reward derivatives and trajectory backpropagation, a substantive improvement for diffusion RL. Joint and single-reward comparisons against DiffusionN
In briefNL2SQL with Iterative Self-Refinement: a system that turns plain-language questions into database queries and checks its own work
SummaryThis paper presents a multi-agent, persona-driven framework that converts natural-language questions into SQL queries, aimed at non-technical users in domains like insurance. The system includes domain relevance detection, schema-grounded knowledge building, KPI-aware SQL generation, syntax self-repair using SQLite execution feedback, and semantic checking by translating SQL back into natural language. Tested on 60 queries (50 from a HuggingFace text-to-SQL dataset, 10 custom domain queries), it reached 90% correctness using Mistral across seven coordinated LLM calls.
In briefANCHOR: A benchmark testing whether AI video-and-audio reasoning is backed by visible evidence.
SummaryThis paper introduces ANCHOR, a benchmark of 2,319 videos and 8,272 question-answer pairs with structured reasoning traces, temporal evidence, bounding boxes, and audio cues, including a human-refined test set. It finds that multimodal models with similar answer accuracy differ markedly in whether their explanations are supported by annotated evidence. Removing video sharply lowers accuracy, while removing bounding boxes or audio causes smaller drops.
In briefWhat drives LLM-brain alignment during conversation, in production and comprehension
SummaryThis paper uses fMRI data collected during dialogue to study how large language model layers align with brain regions during language production and comprehension. It also examines links between alignment and the intrinsic dimension of model representations, along with the contributions of interpretable linguistic and communicative features. The authors report that alignment patterns differ between production and comprehension, with strong associations to intrinsic dimension in regions tied to each mode.
In briefBottleneck prompt sidecars for stable illumination conditioning in SSM-based low-light enhancement
SummaryThis paper studies how to safely add illumination priors to state-space-model (SSM) based low-light image enhancement. A lightweight estimator predicts an illumination map, which is converted into prompt tokens and injected via a residual sidecar adapter at the bottleneck, leaving the SSM core unchanged. On LOL-v1, LOLv2-Real and LOLv2-Synthetic, the method improves PSNR over the base model by 0.94, 1.24 and 0.93 dB, with the clearest evidence on real data.
In briefQuery-Adaptive Recurrent Tokenization: visual tokens that adjust to each question
SummaryThis paper introduces QART, a backbone-agnostic module that makes image representations depend on the text query and adjust their token budget to the information needed. Starting from frozen encoder features, it builds a compact representation recurrently and adds tokens until query-relevant visual features are reconstructed with sufficient fidelity. On Visual7W and VQAv2, QART reports 81.96% and 76.37% accuracy using about 29 visual tokens, 20-25x fewer than dense VLM inputs.
Bavesh Balaji, Pranav Gopinath, Pouya Sadeghi et al.
In briefStudy of the tradeoff between robustness and geometric accuracy in adversarial training
SummaryThe paper studies non-parametric binary classification under an adversary, measuring geometric accuracy by the Hausdorff distance to the Bayes classifier and robustness by the adversarial budget ε. The authors prove that robust classifiers lose at most O(√ε) or O(ε) accuracy, depending on the regularity of the underlying Bayes classifier. They also show that high-complexity hypothesis classes can yield robust classifiers with precisely quantified accuracy, with results holding for any norm on Euclidean space and for both infinite and finite data.
WhyProves optimal quantitative bounds on geometric accuracy loss for robust classifiers, refining prior incompatibility results with rigorous proofs covering general norms and finite data.
In briefTesting whether Go language models understand 'atari' and 'sente' moves
SummaryThis paper builds a controlled evaluation set for Go terminology, focusing on 'atari' and its relationship to 'sente'. The authors test LoGos, a Go language model fine-tuned on terminology and commentary, on recognizing these concepts and evaluating moves. The model struggles to recognize changes in the atari relationship and their strategic implications. Its terminology predictions are linked to move preferences, even when those preferences contradict the reference evaluation.
In briefMambaAct: Faster Mamba-based action experts for flow-matching robot control models
SummaryThis paper presents MambaAct, an efficient action module for flow-matching vision-language-action robot manipulation models. It keeps the pretrained visual-language encoder of π0.5 but replaces its transformer action expert with a smaller Mamba-based one, using a lightweight Context Compressor to supply visual-language conditioning. The authors report a 10× smaller action expert, denoising latency reduced by 64.8% on LIBERO and 58.8% in real-world deployment, and a 1.08-point gain in mean LIBERO success.
Khuyen Pham, Tran Nguyen Le, Edoardo Fazzari et al.
In briefLarge diffusion models need more than EMA: balancing variance reduction with continued optimization
SummaryThis paper compares three training strategies for large diffusion models: constant learning rate with exponential moving average (EMA), learning-rate decay, and Schedule-Free optimization. The theory shows that averaging a constant-step trajectory has an inherent optimization limitation, while LR decay and Schedule-Free admit vanishing stationarity guarantees under standard nonconvex assumptions. Experiments show Schedule-Free consistently outperforms constant-LR EMA across training stages, with better generation quality, and its benefits extend to ReFL and generator-side DMD.
In briefTraining language models to ignore irrelevant demographic cues
SummaryThe paper shows that standard alignment methods such as reinforcement learning improve overall response quality but widen quality gaps across demographic groups, amplifying counterfactual bias. The authors propose a policy gradient method with a group-level reward computed on matched pairs of counterfactual prompts, and build a dataset of controlled prompt pairs. Experiments on several fairness benchmarks show reduced counterfactual differences while overall response quality is preserved.
Basar Demir, Emily Shen, Ksheeraj Sai Vepuri et al.
In briefBeyond Surface Cues: Robust Detection of LLM-Generated Korean Text
SummaryThis paper introduces KoLing, a bias-controlled adversarial benchmark of 50.9k instances that removes superficial shortcuts and separately evaluates detection of LLM-generated and LLM-modified Korean text. It also proposes KoLingProbe, an interpretable detector grounded in theoretical linguistics that models Korean-specific properties across multiple linguistic levels. Experiments show it substantially outperforms strong baselines on all detection tasks and remains robust to adversarial attacks.
In briefMany Biases, One Circuit: How Post-Training Adjusts a Pretrained Cue-Routing Mechanism
SummaryLanguage models sometimes abandon a correct answer when the prompt contains a cue pointing to a wrong option, a pattern linked to sycophancy and related biases. The authors trace the attention heads behind these biases across two post-training pipelines. In multiple-choice tests, four cue types largely rely on the same few heads, a circuit already present in pretrained models. Post-training, especially the final reinforcement learning stage, changes when and how strongly models use this circuit.
In briefNormalizing ASR-Style Transcripts Can Remove Cues Needed for Dialogue Act Classification
SummaryThis paper studies how common ASR transcript normalization, such as removing filled pauses and punctuation and expanding colloquial forms, affects dialogue act classification. The authors test nine transcript conditions across six classifiers and three conversational corpora, with multi-seed and real-ASR experiments. They find that heavy normalization generally reduces Macro-F1, and that the effects depend on preprocessing definitions, feature extraction, and conversational context, so normalization should be task-dependent.
In briefConformal Policy Selection: deciding when AI decisions should be deployed for targeted intervention
SummaryThis paper addresses how to deploy AI policies in high-stakes settings such as medicine and finance, where it matters not only whether a policy helps on average but for whom. It treats each candidate individual as a hypothesis test of whether the AI policy's outcome exceeds a user-specified threshold, and controls the false discovery rate among the selected set. The method, Conformal Policy Selection, combines weighted conformal p-values that correct for distribution shift with a two-step selection-and-pruning procedure. The authors prove finite-sample FDR control and a doubly robust asymptotic guarantee, and verify them with synthetic and semi-synthetic experiments.
Junyu Cao, Ruijiang Gao, Maytal Saar-Tsechansky et al.
In briefDICL-Qwen3VL: A vision-language model for explainable work clothing detection with separate confidence scores
SummaryThis paper addresses difficulties in checking whether workers wear compliant clothing at indoor sites, where poor lighting and occlusion make clothing hard to see. The authors propose DICL-Qwen3VL, which uses LoRA-adapted Qwen3-VL to localize upper-body clothing and identify workwear, and separates visual observations from compliance reasoning with distinct confidence scores. Training with GRPO jointly optimizes detection accuracy, output validity and calibration. It achieves mAP@50 of 0.730 and 0.753 on two datasets, outperforming the evaluated baselines, and provides explicit evidence to support manual review.}
In briefMulti-turn jailbreaks need only about half their queries to reach the same progress
SummaryThis paper studies whether multi-turn jailbreak attacks still need their full earlier query history once an effective attack path is found. The authors remove early attacker queries, rerun the remaining ones in a new conversation, and define the effective horizon as the shortest query sequence that recovers the same attack progress. They find that roughly half of the original queries are sufficient on average, which they call effective-horizon compression. The authors also report that this shorter history helps decide which early queries should get less learning signal during attacker training, and that short reward-propagation windows keep defense calibration performance close to full-trajectory returns.
In briefLeMA: A Scalable Levenberg-Marquardt Method for Training Large Neural Networks
SummaryThis paper proposes LeMA, a distributed Levenberg-Marquardt algorithm for training large neural networks on multi-GPU platforms. It uses an exact dual formulation so memory no longer scales with the parameter count, and a Jacobian sharding strategy that computes shards in parallel without building the full Jacobian. The authors report that this makes exact LMA steps feasible for billion-parameter models, and that an 8-GPU node achieved a 7.97x speedup with 99.7% strong scaling efficiency on a convolutional network with about 1 billion parameters.
In briefSep-Latent: Using Separator Tokens as Anchors for Compact Implicit Reasoning
SummaryLarge language models reason well with explicit chain-of-thought, but this is costly and verbose. This paper proposes Sep-Latent, which uses separator tokens as anchors that compress reasoning logic through an Attention-Decay mechanism, with multi-level distillation aligning explicit and implicit pathways. The method adaptively ends latent thinking based on query complexity. On the GSM8k-Aug benchmark it reaches 56.8% accuracy using an average of only 5.4 tokens.
In briefODSS-Net: Decoupling Shared and Specific Features for Multi-Modal Object Re-ID
SummaryThis paper proposes ODSS-Net, a framework for multi-modal object re-identification that uses data from different sensors. It extracts modality-shared structural features, then uses orthogonal projection to learn modality-specific features free of shared components. A mixture-of-experts module uses the shared structure to adaptively fuse these specific features. The authors report that experiments on three benchmarks demonstrate its effectiveness.
In briefMaskMerge: Structure-aware token merging for faster instance segmentation
SummaryToken merging makes Vision Transformers more efficient by combining redundant tokens, but existing methods rely mainly on feature similarity. In instance segmentation, adjacent objects of the same class can have very similar features, causing tokens from different instances to be wrongly fused. MaskMerge uses coarse intermediate mask predictions to constrain feature-based merging and preserve instance boundaries. On PanNuke and iSAID it improves mAP over the ALGM method while reducing computation and running time.
Bavesh Balaji, Letian Jiang, Alexander Wong et al.
In briefWhere Is the Indirect Evidence? Using Influence Functions to Study Grammar Generalization
SummaryThis paper uses influence functions to measure how individual training sentences affect language model predictions across 15 grammatical phenomena from BLiMP. After removing direct evidence with hand-built rules, the authors treat the 60,000 most influential remaining sentences as indirect evidence. Retraining without these sentences lowers accuracy more than removing matched random sentences, and clustering by overlap recovers three of five BLiMP categories.
William Dai, Zhenghao Herbert Zhou, Maya Viswanathan et al.
In briefDynamical Kernel Attention: Inferring the Present from Past Records
SummaryThis paper introduces Dynamical Kernel Attention (DKA), which derives attention from robust estimation of a latent state evolving under stochastic dynamics. As past states decay toward equilibrium, the information they carry about the present vanishes, bounding the total evidence an attention head can use. The authors state that DKA can abstain when relevant evidence is weak and keeps error bounded when old tokens are discarded, while preserving the parallel structure and cost of standard attention. Empirical evaluation is deferred to a later version.
In briefReward Engine: A Reward-Centric Self-Evolution Method for Open-Ended Tool Agents
SummaryThe paper introduces Reward Engine, which builds a reward for open-ended tool agent tasks that lack an executable ground truth. From the task and initial trajectories, it generates a frozen RewardPack of correctness testcases and graded quality rubrics. On WorkBuddy Bench Office and Claw-Eval, the reward largely agrees with benchmark graders, and using it to revise outputs raises the Pass³ consistency metric from 45.0% to 54.0% and from 78.4% to 83.7%.
In briefCOMET: A Method for Task-Agnostic Continual Vision-Language Learning
SummaryThis paper addresses how vision-language models such as CLIP degrade when they must learn tasks sequentially without task labels or past data. The proposed COMET method uses a frozen teacher for feature reference, Tri-Affinity Distillation to preserve image, text, and cross-modal geometry, and a contextual bandit with a reconstruction router to manage per-task experts. The authors report that each component improves continual learning performance on the X-TAIL benchmark.
In briefFederatedBytes: Byte-level small language model for efficient multimodal federated learning
SummaryThis paper introduces FederatedBytes, a small language model that processes any data, including text and images, as raw byte sequences. It uses a linear attention mechanism to reduce memory and computational costs, making federated learning more practical for NLP and multimodal tasks. Experiments on sentiment analysis with the Multilingual Amazon Reviews and ScienceQA datasets report higher accuracy than ByT5 and LSTM, along with significantly lower memory and communication overheads.
Ali Sarabadani, Massimo Piccardi, Mehran Abolhasan et al.
In briefA feature selection layer for interpreting multi-class neural network predictions
SummaryThis paper extends the Feature Selection Layer, an embedded feature weighting method for neural networks, to better capture label-specific feature relevance in multi-class problems. The authors identify the most effective extension through extensive experiments and benchmark it against established feature weighting methods. The proposed method matches existing techniques in identifying globally relevant features and outperforms them on label-specific relevance. It uses the trained network's internal dynamics, so no post-processing beyond training is required.
Jo\~ao Pedro Silveira e Silva, M\'arcio Dorn, Bruno Iochins Grisci
In briefWhen to Forget: Risk-Controlled Context Adaptation for Long-Horizon Agents
SummaryThis paper studies when long-horizon agents should keep, compress, or remove a block of execution history to improve later performance. It introduces RECAP, which combines matched context interventions, execution-relational prediction, and independent policy certification under risk constraints. On Hunyuan-A13B, the authors report editing 43.24% of retained checkpoints while matching the full-history baseline's observed 91.89% continuation success and reducing mean continuation prompt cost by 4,619.8 tokens (2.13%).
In briefHAPACT: A Benchmark for Locating Human-Body Physical Impacts in Movies
SummaryThis paper introduces a new task that locates physical impacts in a video in time and on the human body. The authors release HAPACT, a dataset built from movie clips with over 2,800 video-query pairs and more than 20,000 fine-grained human-annotated frames. They evaluate existing video moment retrieval and human-scene contact models, and propose a text-conditioned spatial baseline. Results show that impact localization remains challenging for current methods.
In briefFusionRare: A Multi-Paradigm AI Framework for Rare Disease Diagnosis
SummaryThis paper presents FusionRare, a framework that combines symbolic knowledge reasoning, deep learning, patient case retrieval, and large language model inference to diagnose rare diseases. Tested on 2,528 strictly isolated clinical cases, it reached a macro-averaged Recall@1 of 51% using phenotypes alone and 79% using phenotype-genotype analysis. A randomized clinician-AI study reported improved true-diagnosis ranking in 45% of decisions and a 21% reduction in time to diagnosis.
WhyFuses four diagnostic paradigms, validated on 2,528 isolated cases plus a randomized clinician crossover study, giving solid evidence.
In briefTesting whether language models reliably apply logical reasoning skills across problem families
SummaryThe paper argues that current benchmarks report single accuracy scores that say little about whether models generalize their reasoning. It proposes competence-based evaluation (CBE), which scales problem complexity while keeping the required skill fixed and checks answers across equivalent rephrasings. Tested on ordering tasks with six GPT and Qwen models, performance was reasonably good over bounded complexity ranges, but even strong models failed the competence condition across the tested family.
In briefBengali Hallucination Detection Using Retrieval and Programs: A System Study
SummaryThis paper presents an inference-only system for detecting hallucinations in Bengali text. It combines query rewriting, hybrid retrieval, and task-specific language-model judges, routing each example to either a 27B textual judge or a 14B agent that can compare answers, reason, or run Python code. The authors analyze the system's code, saved predictions, and execution traces from the Olikbochon challenge. They also report overlap between corpus records and evaluation inputs, so they do not claim unqualified generalization beyond the corpus.
In briefWhat Accuracy Hides in Learned Transformer Programs: Inspecting Extracted Code
SummaryThis paper studies Transformer Programs, Transformers constrained so their learned computation can be extracted as readable Python. Across nine algorithmic tasks, the authors compare token accuracy with whole-sequence accuracy and find failures that sampled tests miss. Training data revised using program inspection and failure analysis improved sequence accuracy on several tasks.
In briefNLData2Opt: A Benchmark for Deriving and Predicting Optimization Parameters from Language and Data
SummaryThis paper introduces NLData2Opt, a benchmark that tests whether large language models can build optimization models from problem descriptions and raw data. Models must extract relevant parameters from data containing misleading attributes, compute missing parameters from existing data, and predict unavailable future values from historical observations. Results show substantially lower performance than on existing benchmarks, indicating that progress on those benchmarks does not directly carry over to formulation accuracy in practical optimization workflows.
In briefBatched Contextual RL II: Can RLVR Train With Shorter Rollout Windows?
SummaryThe paper studies whether reinforcement learning with verifiable rewards (RLVR) needs long rollout windows for math reasoning. It proposes Concurrent Constraint Reinforcement (CCR), which places three verifiable problems in one capped completion so correct answers finished before truncation still earn credit. With an 8,000-token window, CCR reduces tokens per training step by 82.1% on 4B and 79.1% on 8B compared with single-problem GRPO at 32k.
In briefLearning to Be Guided: Shaping Latent Geometry for Guidance-Compatible Diffusion
SummaryThis paper studies how class arrangement in a latent space affects classifier-free guidance (CFG) in conditional diffusion models. Under a Gaussian class model, the authors propose Prototype Regularization (PR), a lightweight surrogate objective applied to pooled latents. On ImageNet, adding PR to a latent diffusion baseline lowers guided gFID from 3.31 to 2.28, and the method also generalizes to clustering pseudo-labels, text-embedding conditions, and video latents.
In briefToPA: Block-wise Toeplitz Adaptation for Efficient Fine-Tuning of Large Models
SummaryThis paper proposes ToPA, a parameter-efficient fine-tuning method that updates pre-trained model weights using products of block-wise Toeplitz matrices, without explicit low-rank or sparsity constraints. The authors theoretically show that products of block-wise Toeplitz matrices can approximate arbitrary matrices with shorter chain lengths, and that ToPA is more expressive than existing low-rank and sparse parameterizations. Experiments across 21 NLP and CV datasets and 5 model architectures are reported to validate its effectiveness.
In briefCooperation Without Adaptation: Testing LLM Agents in a Collective-Risk Game
SummaryThis paper tests whether five language-model agents adjust their contributions to a shared pool in a collective-risk dilemma as risk and partner behaviour change, or follow a fixed strategy. Most models contributed similar shares at every risk level, and most still contributed when risk was zero. Several models could identify the higher-payoff action yet changed their behaviour little. The authors conclude that cooperative self-play alone does not show strategic adaptability.
Trung-Kiet Huynh, Dao-Sy Duy-Minh, Chi-Nguyen Tran et al.
In briefProtoGTX: Prototype-aligned graph transformer that explains whole-slide pathology image predictions
SummaryProtoGTX is a graph-based method for analyzing whole-slide pathology images that aims to explain model predictions. It first contextualizes patch features over a spatial graph, then learns morphological prototypes via cross-attention and aligns them with a fixed reference prototype vocabulary. On the CPTAC and BRACS datasets it outperforms the compared baselines, and its prototype maps co-localize with pathologist annotations on BRACS.
In briefLooking Back to Correct Forecasts: Error Feedback Calibration for Time Series
SummaryThe paper proposes Error Feedback Calibration (EFC), a test-time method that post-hoc corrects a frozen time series forecaster using errors revealed by past overlapping forecasts. It combines a Causal Error Diagonal, Cumulative Error Transport, temporal out-of-sample validation, and Temporal Residual Learning. The authors report that EFC broadly improves forecasting performance across backbones and datasets while substantially limiting harmful over-correction.
In briefShadowFPT: A backdoor attack on federated prompt tuning using hidden shadow triggers
SummaryThis paper examines a security risk in Federated Prompt Tuning, where a frozen shared backbone lets a malicious client implant backdoors while keeping its uploaded prompts similar to benign updates. The proposed ShadowFPT attack pretrains a learnable Shadow Trigger against the frozen CLIP visual encoder, steering triggered inputs toward a target class in representation space. In the main setting, attack success rises from 19.81% to 90.36% while clean accuracy is maintained. The authors conclude that defenses relying only on prompt-update anomaly detection are insufficient.
In briefMS-Eval: A benchmark for evaluating sub-agent delegation and orchestration in tool-heavy LLM agent tasks
SummaryThe paper introduces MS-Eval, an evaluation kit for the Main-Agent Delegation and Sub-Agent Execution (MAD-SAE) paradigm, in which LLM agents hand sub-tasks to specialized sub-agents. It includes an MCP-intensive benchmark, a controlled execution harness, delegation and orchestration metrics, and the MSQ index. Across four MAD-SAE variants and eight frontier models, sub-agent specialization performs best, open-source models are largely competitive, and MAD-SAE generally raises cost with benefits that depend on the model and task.
In briefEvo-B&B: LLM-guided joint evolution of branching and node selection policies for MILP solving
SummaryThis paper proposes Evo-B&B, a framework that uses a large language model to guide the evolution of executable branching and node selection policies for mixed-integer linear programming in branch-and-bound. It uses end-to-end feedback from the SCIP solver and a behavior-aware multi-island mechanism that measures policy similarity on fixed decision states to keep the population diverse. Experiments on four standard MILP benchmarks across multiple instance scales report competitive solving performance and good generalization to larger instances. Ablation studies support the benefits of joint evolution and behavior-aware diversity management.
In briefUncertainty-aware budget allocation for efficient LLM tree search
SummaryThis paper addresses tree search for multi-step reasoning in LLMs, where existing methods expand every step with a fixed budget regardless of marginal benefit. The authors find that continuations sampled from a shared state are largely duplicates, and that the semantic dispersion of sibling actions predicts their downstream redundancy. They propose UAA, a training-free framework that uses values to pick promising candidates and sibling-dispersion uncertainty to decide how many to expand. Across math, logic, and planning benchmarks, UAA reports up to 90% fewer tokens and up to 9x lower latency at comparable accuracy.
In briefAnchorQ: Sample-Efficient Robot Policy Improvement via Flow-Reversed Action Anchors
SummaryThis paper studies how reinforcement learning can improve generalist robot policies while reducing costly online interaction. It proposes AnchorQ, which uses flow reversal to map reference behaviors to state-conditioned latent anchors, then samples local particles and selects refinements guided by a value function. The authors report that under the same interaction budget, AnchorQ outperforms state-of-the-art post-training baselines, with a 37% gain in real-world success rate and 97% higher simulated task throughput.
In briefAdaptive light-cone transform for more robust non-line-of-sight imaging
SummaryThis paper proposes the Measurement-Adaptive Light-Cone Transform (MALCT), which keeps the efficient Fourier-domain inversion of the standard light-cone transform while combining analytical and learned responses based on measurement conditions. An attention-enhanced spatio-temporal module is added to support reconstruction under low-signal conditions. Experiments on synthetic and real NLOS measurements show improved reconstruction quality, with robustness assessed under temporal broadening and reduced acquisition time.
In briefBenchmarking and improving LLM agents for causal discovery
SummaryThe paper argues that current benchmarks let LLMs guess causal directions from meaningful variable names, so they cannot show whether agents can inspect data and correct algorithm errors. The authors introduce Heuresis, a benchmark of anonymous simulated datasets with mechanisms ranging from linear Gaussian to heavy-tailed, which tracks every edge an agent changes. Results report that agents using the data surpass the best chosen algorithm on all six dataset families.
In briefDinoMUSE: DINOv2-guided method for estimating sideways tissue motion in ultrasound elastography
SummaryThis paper addresses poor accuracy in lateral motion estimation for ultrasound elastography. It proposes DinoMUSE, which uses a mixture-of-experts module to fuse DINOv2 visual features with local matching features. On simulation data it reduces lateral displacement and strain mean square error by 60% and 65% versus a state-of-the-art method, and it improves lateral strain SNR and CNR on in-vivo breast data while keeping axial performance competitive.
In briefA method to reduce bias from combined vision and language cues in visual question answering
SummaryVisual question answering models tend to overfit frequent patterns, a problem known as bias. The authors study multimodal bias separately for counting and non-counting questions, finding that each comes from different sources: counting bias from the vision side propagated through fusion, and non-counting bias mainly from language. They propose VDB, which uses variational bias models based on Variational Auto-Encoder principles plus a bimodal bias model. Experiments on VQACPv2, VQACPv1, and GQAOOD datasets support the claims and the method's effectiveness.
In briefVTrajEmbed: Matching Text to Video Object Paths Using Space-Time Embeddings
SummaryThis paper presents VTrajEmbed, a model that splits a video into object trajectories and matches each trajectory's embedding to text, enabling object-level retrieval. It uses a dual-branch Slow-Fast encoder, mask-guided contrastive learning, and a new trajectory-level benchmark called Text2Traj-Bench. According to the paper, VTrajEmbed-5B improves text-to-trajectory and trajectory-to-text R@1 over Qwen3-VL-Embedding-8B by 15.48 and 18.05 points, and reaches 49.48% average R@1 across five global retrieval benchmarks versus 42.86%.
In briefReliance Hypothesis: Lower-resource languages make models' reasoning chains more faithful
SummaryThis paper examines whether large language models' chain-of-thought (CoT) reasoning faithfully reflects how they reach answers. Using parallel multiple-choice questions from MMMLU in eight languages and four models, the authors compare language resource level with task type, knowledge-intensive versus reasoning-intensive. Faithfulness rises as resource level falls on knowledge-intensive tasks but not on reasoning-intensive ones. The authors propose the Reliance Hypothesis: models lean more on their reasoning when they cannot recall an answer.
WhyLinks CoT faithfulness to language resource level and task type, proposing a Reliance Hypothesis supported across eight languages, four models, and a Simpson's paradox.
In briefToo Emotional to Be Fake? Emotion Shifts Within Speech Expose Deepfake Detector Weaknesses
SummaryThis paper introduces EmoShift, a benchmark comparing speech with a stable emotion against speech whose emotion changes within one utterance, to test how emotion dynamics affect speech deepfake detection. It includes 7,240 bona fide utterances and 9,720 synthetic utterances from four expressive TTS systems. Across seven detectors, emotion changes made synthetic speech harder to flag, raising spoof miss rates by up to 25.3 percentage points, while changing-emotion genuine speech raised false alarms by up to 24.4 points.
In briefPredicting how a dialogue turn's annotator panel votes from partial earlier votes
SummaryThe paper studies partial-revelation panel forecasting, predicting the full count distribution of the next dialogue turn's annotator panel given k of n votes on the previous turn. It defines coarsening-consistent panel operators with a closed-form Dirichlet-multinomial realization and compares them with per-level lookup on IEMOCAP, M3ED, and EmoWOZ. The reported gains are modest and corpus-dependent, with several partial-revelation differences being small or null.
In briefWorld Action Critic (WAC) Model: a method that learns to predict future outcomes of candidate actions and to judge how good they are, for training decision-making policies.
SummaryThe paper proposes the World Action Critic (WAC) Model, which generates multimodal action candidates and jointly learns candidate-conditioned future prediction and consequence evaluation. Critic signals shape intermediate predictive representations during rollout. Applied to autonomous driving, WAC reaches 93.4 PDMS on NAVSIM-v1.
In briefVISTA: A Benchmark for Goal-Directed Vision Assistance for Blind and Low-Vision Users
SummaryThis paper introduces VISTA, a multimodal egocentric benchmark for testing AI assistance that helps blind and low-vision users complete everyday tasks. It contains about 1,000 recordings captured with Project Aria glasses, with synchronized sensor streams, annotations, and ten task categories. The authors evaluate 14 vision-language models zero-shot and provide a lightweight baseline that integrates all six modalities. Results show that existing models can often describe scenes but remain limited in generating goal-directed assistive guidance.
No\'e Mosca, Mohamed Rayan Barhdadi, Muhammad Hamza et al.
In briefCausal Edit Serialization: a way to express local revisions within one autoregressive stream
SummaryThis paper introduces Causal Edit Serialization (CES), a framework that lets standard autoregressive language models express revisions such as insertion, deletion, replacement, copying and movement within the same causal token stream, using pointer tokens paired with predicted source addresses. It uses a lightweight localization head and dual-index rotary position encoding while keeping a standard causal attention mask, and is pretrained on ordinary language modeling combined with serialized repairs. The authors report that CES models complete applied edits with modestly higher clean-text bits per byte than clean autoregressive models, and that a grammatical-error-correction proof of concept produces more beneficial than harmful changes under reference-blind model judging.
In briefPruning by replay and integer-only inference for fast, compact speed-sign recognition in cars
SummaryThis paper presents a two-stage speed-limit sign recognizer for automotive camera chips. The detector is compressed by replaying channel deletions on the original network followed by distillation recovery, while the classifier and detector post-processing use integer arithmetic only. On YOLOv5n, replay removes 60.3% of parameters while matching the parent's accuracy. The combined 0.72M-parameter system recognizes 95.2% of signs from a held-out automotive camera at 90% precision and 94 frames per second.
In briefA comprehensive study of how the position of answer choices and documents affects large language models.
SummaryThis paper examines position bias, where the placement of answer options or documents influences the outputs of large language models. The study covers models ranging from billions to trillions of parameters, across multiple-choice and multi-document tasks in five languages. It finds the bias is noticeable in smaller models and non-English languages, but is significantly reduced by increasing model scale and using standard Chain-of-Thought prompting. The authors also report that emerging thinking models with extended reasoning are nearly immune to it.
In briefCompression Precedes, Content Decides: Emergent Misalignment in Reward Models
SummaryThis paper fine-tunes two Skywork-Reward-V2 reward models to prefer harmful responses in either medical or financial domains, and finds this preference generalizes to seven held-out domains. The fine-tuned models drop only 4.8 and 5.3 points on RewardBench 2 and still outperform the unmodified Qwen reward model. The authors report that reward models are more prone to preference inversion than their language model counterparts, and localize the mechanism to the last scored token: harmful and safe representations are geometrically compressed, and harmful representations align more closely with the scalar head, earning higher rewards.
In briefCLARO: A VLM agent that coordinates perception and manipulation to grasp language-specified objects in clutter.
SummaryThe paper presents CLARO, a vision-language model agent that orchestrates perception and manipulation skills to locate and grasp a target object described in natural language within cluttered scenes. It keeps a memory of past actions to support long-horizon execution and outputs grasp poses or 3D waypoints that transfer across robot arms without task-specific training. The authors report an average success rate of 89.1% in simulation and 88.8% when deployed on a real robot arm.
Sungsu Kim, Vineet Bhat, Prashanth Krishnamurthy et al.
In briefReasoning reranking that ranks items using only Semantic IDs
SummaryThis paper studies whether reasoning rerankers can rank a candidate list using only compact Semantic IDs, without reading item text. The authors build ReRankID, a controlled testbed, and find that untuned models often copy the input order instead of reranking. Post-training methods, especially an 8B model trained with a graded ranking reward, fix much of this and match text-reading LLMs with much longer prompts, but describe items less accurately.
In briefCausal Latent Workspaces for Looped Transformers
SummaryThe paper proposes a causal latent workspace Transformer, a looped model that reuses one shared block across four rounds. Each position writes gated vectors into prefix-specific latent slots and retrieves a content-addressed summary, adding to causal self-attention. Across three byte-level corpora, three seeds and two parameter budgets, it achieves the best mean held-out perplexity in five of six settings. Adding the workspace improves over the anchored loop in all six settings.
In briefAnalyzing Binary O-type Stars with Neural Networks
SummaryThis paper uses a recurrent neural network to estimate effective temperature and surface gravity for simulated O-type binary systems, using 34 equivalent-width measurements from optical spectra. Secondary spectra were added to primary spectra at different contributions. For most spectra, estimated values decreased as the secondary contribution grew, but behavior varied across combinations. The authors state the system is sensitive to the added secondary but does not determine a unique secondary contribution for observed binaries.
Miguel Flores R., Luis J. Corral, Celia R. Fierro-Santill\'an et al.
In briefSTRIDE: Testing AI Driving Models' Space-Time Reasoning in Tough Edge Cases
SummaryThis paper introduces STRIDE, a benchmark for testing vision-language models on safety-critical, rare driving scenarios, covering six task families such as spatial perception, temporal memory, and trajectory prediction. It combines multiple-choice questions with free-form visual question answering. The best model, GPT-6-Astra, reaches only 43.4% average multiple-choice accuracy, compared with about 20% for random guessing, and the authors find that describing scenes well does not guarantee accurate safety judgments.
In briefOri: Drift-Aware Streaming k-PCA Optimizer for Very Low-Rank LLM Training
SummaryThis paper proposes Ori, a low-rank optimizer for training large language models that tracks how gradient principal directions slowly shift over time. It uses a streaming k-PCA method inspired by Oja++, whose low per-step cost allows frequent on-the-fly subspace updates. The authors report that on 1B LLaMA pretraining Ori narrows the gap to full-rank optimizers by up to 69.3%, and it performs best on 10 of 16 scenarios in Roberta-Base fine-tuning.
In briefPruning Leaky Neurons: Reducing Private Data Leakage in Fine-Tuned LLMs with Joint-Embedding-Guided Pruning
SummaryThis paper addresses the risk that large language models fine-tuned on private records leak personally identifiable information. It proposes JEPA-Prune, which uses a public model trained only on PII-free views of the data as a reference to identify PII-responsive neurons via integrated gradients and a specificity score, then prunes them without retraining. Experiments report that pruning about 1.6% of neurons sharply reduces PII inference and extraction, while perplexity on unseen text stays within 16% of the unpruned model for most LLMs.
In briefSelf-Evolving Coding Agents via Harness-Policy Optimization: a method where coding agents improve both their task harness and their policy in alternating rounds.
SummaryThe paper proposes Self-HPO, a recursive self-improving framework that alternates optimization of a coding agent's harness and its policy model. The policy first analyzes its failures and proposes harness candidates, then is updated with reinforcement learning under the selected harness, with hard training tasks guiding the next harness search. Across three coding benchmarks, the authors report it outperforms harness-only and policy-only baselines, raising Qwen3.5-35B-A3B's average resolve rate from 57.8% to 64.8%.
In briefNext Token or Vector? Learning How Far to Decode
SummaryThe paper proposes CADENCE (Context-Adaptive DEcoding), a generation method that predicts several future tokens at once using Gaussian prediction heads and a shared decoder. A learned router chooses the prediction length from the model's hidden states, falling back to single-token generation when needed. Experiments with Llama-3.1-8B-Instruct report up to 1.3x higher throughput than standard greedy decoding.
Sai Sreenivas Chintha, Chidera Biringa, Rishab Sharma et al.
In briefAnytime Contracts for AI Services: contract design to curb hidden quality downgrades when users can cancel at any time
SummaryAI service providers may secretly route requests to cheaper, lower-quality models, and the randomness of model outputs makes this hard to detect. The paper extends contract theory to settings where users can terminate at any time, formalizing the problem as a Principal-Agent game. It shows that budget-optimal contracts are proportional to anytime-valid hypothesis tests based on likelihood-ratio e-values, and gives an efficient algorithm and an e-process approximation. The approach is evaluated on synthetic data and data from deployed LLMs.
In briefWhen Can an Agent's Self-Verifier Be Trusted? Stress-Testing Rationale Verification Under Optimization Pressure
SummaryThis paper examines whether an AI agent's check of its own stated design rationales stays reliable once the agent optimizes against it. The authors propose a pressure ratio and four probes to stress-test such verifiers, and apply them to their own Rationale Consistency Score within an LLM agent designing machine-learning pipelines. They report that the verifier barely shaped search in their prior setup, and that when the agent could query it before committing, claims passing the check rose while held-out faithfulness fell. They recommend auditing verifiers with these probes before trusting them.
In briefDirect conditional parameters for N-dimensional splatting, replacing covariance-derived readouts to speed up rendering
SummaryThis paper addresses the conditional slicing step in N-dimensional splatting, which extends 3D Gaussian Splatting with view direction and time to model view-dependent and dynamic effects. It replaces covariance-derived readouts with explicit per-effect parameters, applied to both Gaussian (dGS) and Beta (dBS) kernels. Across five static and dynamic benchmarks, the method matches or exceeds covariance-derived versions on quality, with gains up to +1.26 dB in the main MCMC evaluation. Production slicing is 6.9 to 7.7 times faster, and complete-render speedups range from 1.25 to 1.91 times on four trained scenes.
Zhongpai Gao, Benjamin Planche, Van Nguyen Nguyen et al.
In briefTransferring calibration gains between model representations
SummaryThis paper studies how much of a calibration update learned for one representation of a pretrained model can be reused by another representation of that model. The method extracts positive group-wise gains and installs them into the recipient's weights or native quantization scales without recipient-side optimization. For a Qwen3-1.7B floating-point and GPTQ INT3 pair, bidirectional transfer retained 76.9% to 90.5% of the text-loss improvement on fresh WikiText and C4 data.
In briefCombining unit-level estimates breaks coverage of group-effect intervals in causal foundation models
SummaryCausal foundation models estimate treatment effects one unit at a time, but many decisions rely on group-averaged effects. The authors find that credible intervals for such averages under-cover on synthetic data and on ACIC 2016, because unit-level variances are combined as if errors were independent while posterior uncertainty makes them correlated. Two neural process models that predict jointly for a group are shown to account for this correlation and improve coverage across the benchmarks.
In briefOne Communicator, Three Vendors: Extending MoA Predictive Theory to a Real NVIDIA+AMD+CPU MPI Ensemble
SummaryThis paper extends earlier MPI work to a communicator spanning multiple hardware vendors, placing A100 and MI100 GPUs and CPU ranks under one Slurm heterogeneous job. It gives a weight-split theory and reports measurements, including that MI100 time is not linear in rows and that changing launch settings cut its per-call energy by 32x. The authors also show that the earlier companion paper's 8.1x A100 speedup and 22 ms Gatherv result were artifacts of several ranks sharing one GPU.
In briefStoBasket: Optimizing Upsell and Cross-Sell Together for Contribution Margin
SummaryThis paper proposes joint stochastic basket editing, a method that selects both replacement offers (upsells) and addition offers (cross-sells) for an existing shopping basket. Offers are valued by expected incremental contribution margin, with upsell and cross-sell responses coupled. In simulations calibrated on five public basket datasets, the approach captures nearly all attainable margin and outperforms baselines such as margin-weighted popularity and a next-basket recommender.
In briefBayesian optimization over discrete spaces using acquisition-guided generation (AFGG)
SummaryThis paper proposes acquisition function guided generation (AFGG), a Bayesian optimization method for discrete design spaces. It pairs a pretrained generative model with a surrogate defined directly in data space, sampling candidates from a distribution tilted by the acquisition function via a diffusion bridge. The method is instantiated with a categorical diffusion model and an autoregressive normalizing flow. On a range of molecular benchmarks, the diffusion variant performs comparably to leading methods, while the flow variant outperforms them in most cases.
Tommaso Ancilli, Thomas Wedenig, Henry Moss et al.
In briefShape-Aware Neural Time Fields for Rigid-Body Motion and Manipulation Planning
SummaryThis paper proposes Shape-Aware Neural Time Fields, a physics-informed neural framework for planning rigid-body motion. It learns a reusable travel-time (cost-to-go) field that separately captures translation and rotation constraints from surrounding geometry, paired with a sampling-based planner. Evaluations on 2D and 3D rigid-body planning and object pushing show higher success and faster inference than physics-informed neural baselines, with lower planning times than classical samplers. Closed-loop pushing experiments show fast replanning supports adaptive execution under deviations and disturbances.
In briefBalancing Spoilers and Guesswork in Coding Benchmark Task Specifications
SummaryThe paper examines two problems in AI coding benchmarks: under-specified tasks that force guesswork or contamination, and over-specified tasks that leak answers and skip intended challenges. It shows that some benchmarks, including SWE-bench Pro, FEA-bench, Doc2Feat-bench, and Multi-SWE-bench, add ground-truth-derived hints that cause over-specification. The authors test different specification levels and recommend methods for deriving parsimonious hints, and they plan to release their hints.
Yuanzhe Liu, Bastian Lipka, Avraham Shinnar et al.
In briefCorrecting Input and Feature Differences for Cross-Hospital IHC Grading
SummaryPathology foundation models are mostly trained on H&E images but are often used as frozen feature extractors for immunohistochemistry (IHC), where differences between hospitals can hurt grading accuracy. The authors benchmark label-free correction methods applied to input images and to extracted features, using UNI and Virchow2 across three IHC markers and six cross-hospital transfers. Their proposed FIX method gave the largest macro-F1 gain among input-space methods, while simple feature-wise standardization performed best in feature space. Combining both approaches yielded complementary improvements.
In briefRefusal Is Not Robustness: Measuring Hidden Harmful Reachability in Language Models
SummaryThis paper argues that attack success rate alone is insufficient for judging LLM safety, since it only evaluates sampled outputs. The authors propose bounded continuation analysis (BCA), which treats autoregressive generation as a stochastic process and computes the finite-horizon probability of reaching a harmful continuation. Experiments on four instruction-tuned LLMs and 100 JailbreakBench behaviors show that BCA quantifies hidden harmful behavior and can guide red-teaming efforts.
In briefMedSTP: A Training-Free Token Pruning Method for Medical Visual QA That Preserves Key Visual Evidence
SummaryMedical vision-language models answering visual questions process many visual tokens, and pruning them can risk losing boundary and texture details. The paper proposes MedSTP, a training-free pre-decoding pruning method that keeps retained tokens close to the dominant visual evidence subspace and adds a bounded token-grid transition prior to favor boundary and texture sensitive tokens. It is evaluated on SLAKE, OmniMedVQA, and RadImageNet-VQA using Lingshu-7B and Qwen3-VL-8B backbones. The authors report consistently strong performance on both closed-ended and open-ended RadImageNet-VQA evaluations.
In briefNative geometry versus linear recoverability: how alignment shifts platonic scaling
SummaryThis paper tests whether larger models learn increasingly similar representations, as the Platonic Representation Hypothesis predicts. Using matched Legacy Survey and HSC embeddings across five model families, it compares native cosine neighbourhood structure with correspondence revealed by a supervised Ridge alignment. Ridge alignment strengthens the link between model size and cross-dataset neighbourhood correspondence, and the effect persists after reducing dimensionality. However, the learned maps are far from rotations or uniform scaling, and sparse features offer no advantage, so scale mainly improves linear recoverability rather than establishing convergence to a shared native geometry.
In briefHow generating several people together can change the faces of people not targeted
SummaryThis study examines how, in multi-person personalized image generation, people outside the intended target can also see their facial identity change. It analyzes the UniPortrait system without proposing a new preservation method. Using 119 two-person baseline scenes and InsightFace similarity measurements, it compares single-identity and native dual-identity generation. Native dual-ID generation shows positive similarity changes for both the protected and target persons, but the authors state these results apply only to the evaluated implementation and cohort, not to causal component effects or general identity preservation.
Edrees Al-Edreesi, Ghaleb Al-Gaphari, Muneer Alsurori et al.
In briefOmniXRef: A unified framework for editing videos using mixed text, image, and video references
SummaryOmniXRef is a unified framework that lets users combine any number of text, image, and video references to control various attributes of a source video. Its model, OmniXRef-Model, uses a single-stream diffusion transformer trained on synthesized data with curriculum learning. The authors also introduce OmniXRef-Bench, a benchmark covering camera, motion, content, style, and visual effects. The paper reports that the model outperforms task-specific baselines and commercial models.
In briefTWIN-MoE: A Two-Stage Expert-Based Imputation Network for Incomplete Multimodal Learning
SummaryThis paper addresses missing modalities in multimodal learning, where existing Mixture-of-Experts (MoE) methods either scale poorly or require costly retrieval. The authors propose TWIN, a two-stage framework: pre-imputation generates an estimate in the encoder embedding space, constrained by statistics from observed samples, and post-imputation refines the missing representation by aggregating expert-conditioned observed representations within the same sample. The authors report state-of-the-art results on four missing-modality benchmarks under natural, random, and asymmetric dropout protocols, using substantially fewer parameters than competing MoE-based imputation methods.
Liang Nathan Zheng, Jiyuan Liu, Wenhao Liang et al.
In briefR²Orch: Self-Improving Agent Orchestration via Self-Play and Skill Withdrawal
SummaryThe paper proposes R²Orch, a recursive self-play framework for agent orchestration, with an inner self-play loop that improves the orchestrator and an outer loop that improves the inner loop. It introduces Recursive Learnability, a training objective rewarding task proposers for choosing tasks whose outcomes still differ when skills are withdrawn, and Skill Withdrawal, which removes a skill once controlled evaluation shows the orchestrator performs as well without it. Experiments on twelve datasets across question answering, math, code, and interactive tasks report consistent gains over baselines.
WhyProposes a novel recursive self-play orchestration paradigm with skill withdrawal; evaluated on twelve datasets, though breadth of comparisons is unclear.
In briefSPADE-KD: A Decentralized Distillation Framework for Collaborative Learning Across Heterogeneous Clients
SummarySPADE-KD is a fully decentralized model distillation framework that lets clients with different model architectures and data distributions learn collaboratively, aiming for personalized rather than global performance. It combines adaptive distillation, which weights incoming models by their advantage on local data, with utility-driven peer selection that ranks partners by how much they improve the local model. In federated classification experiments, the authors report lower loss and faster convergence than naive decentralized distillation, and that the two components work better together than separately.
In briefGraph Serializations for Reasoning LLMs: How Text Formats Affect Knowledge Extraction and Graph Reasoning
SummaryThis paper examines how the text format used to serialize a knowledge graph for an LLM affects model performance. The authors evaluate 15 graph representation formats across nine frontier LLM configurations on ontology-based knowledge extraction and structural graph reasoning tasks. Knowledge extraction is relatively insensitive to format, while the best format for structural reasoning significantly affects accuracy. The authors propose treating graph serialization as a dynamic orchestration decision, since the optimal choice depends on the model, query, and task.
In briefThunder-DeID+: A Korean De-identification Model Extended from Court Judgments to Broader Sensitive Documents
SummaryThis paper tests whether a Korean de-identification model built for court judgments transfers to a wider range of sensitive documents. It introduces Thunder-DeID+, which pairs a refined three-tier PII taxonomy with a boundary-aware tokenizer to fix BPE tokens that merge personal information with Korean particles. On full judgments, the tokenizer raises strict typed F1 from 0.76–0.77 to 0.87–0.88, mainly through higher precision. After fine-tuning on 1,000 broader sensitive documents, the model reaches 0.93 F1, indicating effective transfer.
