In briefMathematical Collective Unconscious: Jung's Archetypes in LLM Latent Space
SummaryThis paper asks whether large language models spontaneously form archetype-like regions in their high-dimensional latent spaces, echoing Jung's idea of a collective unconscious. Using OpenAI's text-embedding-3-small model and cosine similarity and angular distance metrics, the authors ran thirteen experiments on synthetic narratives, literature, dream reports, and folklore. The results indicate that archetypal roles occupy distinct, compact regions shaped by narrative structure, and that text category affects spatial distribution, with dream reports concentrating in the Trickster and Shadow regions.
In briefHow Language Models Handle Conflicting and Combined Knowledge During Continual Learning
SummaryThis paper studies how a language model's base knowledge, a parameter update (adapter), and input context interact during continual learning. Using a hypernetwork framework that generates adapters from context, the authors find that under knowledge conflict the model tends to follow the context, with its choice depending more on where knowledge appears than on its content. They also report that models fail to reliably compose knowledge across all three sources in multi-hop reasoning.
In briefFixing low-bit failures in robot vision-language-action models with quantization-conditioned reinforcement learning
SummaryThe paper studies compressing vision-language-action robot policies through quantization, using 4,050 real-world trials across five models. It finds quantization-aware training beats post-training quantization but still falls short of full precision on long tasks and under visual or semantic shifts. The proposed QCRL method is an offline reinforcement learning approach that uses progress-based advantage estimation and quantization-aware conditioning. With INT4 weights on π0.5, QCRL reaches 99.3% of full-precision mean score across three visual-shift conditions and saves 42.9% model memory.
In briefProtoPur: Prototype-Purity Pseudo-Labeling for Weakly Supervised Pathology Instance Learning
SummaryThis paper addresses pathology whole-slide image analysis, where only slide-level labels are easy to obtain and region-level annotations are costly. It proposes ProtoPur, which builds prototype regions from frozen foundation-model features, measures class-balanced local purity, and assigns pseudo-labels to high-purity regions linked to a target class. The authors provide a PAC-style analysis and report experiments on three public datasets with two pathology encoders, showing improved instance-level performance across several MIL backbones while keeping competitive bag-level results.
In briefEvaluating text by masking tokens at different rates in diffusion language models
SummaryThis paper proposes rate-resolved text evaluation, which uses the masking rate of masked diffusion language models as a diagnostic axis. Reconstructing text under different masked-token fractions turns one diffusion score into a vector of band-conditioned scores that reflect reconstruction difficulty at varying visible-context levels. Experiments on SummEval and WMT MQM zh-en show that keeping this rate structure adds complementary signal and can localize error spans. The authors note that band semantics depend on the model and task, so each application requires empirical validation.
In briefA Persian dataset and benchmark for turning ambiguous or infeasible instructions into robot action plans from images
SummaryThe paper introduces the first Persian vision-language-action dataset, with 258 instances over 230 real household and office photos, annotated with reference plans, ambiguity and infeasibility labels, and clarifying question-answer pairs. The authors ablate retrieval augmentation, chain-of-thought prompting, and an agentic disambiguation step, evaluating them with an LLM-as-judge protocol and a new metric, PGOS-FA. Retrieval augmentation shows a large, significant effect, while finer additions are not statistically distinguishable at this scale.
Abdollah Pourasghari, Saeedeh Sadat Sadidpour, Mohammad Roustaei
In briefDrift, Then Spread: One-Step Generative Policies for Online RL
SummaryThis paper introduces DriftSpread, a one-step generative policy for online maximum-entropy reinforcement learning, built on the Drifting framework. It learns a critic-guided Drift update and an entropy-induced Spread update sequentially in the same actor, without iterative sampling or policy-density estimation. The authors prove convergence to the Boltzmann target under exact fitting with a fixed critic. On MuJoCo, it reports the highest aggregate normalized return among evaluated methods and better training and inference efficiency than diffusion- and flow-based baselines.
In briefFeedback as an Interface: Controlling Form and Detail in LLM Translation Refinement
SummaryThis paper studies how the presentation of evaluator feedback affects LLM translation refinement. The authors keep the score, main issue, and recommendation fixed while varying the feedback's form and supporting detail. Across seven models and six translation directions, compact low-load feedback outperformed high-load feedback under xCOMET, limiting quality loss during refinement.
Yiqi Liu, Flavjo Tomori, Nafise Sadat Moosavi et al.
In briefSparse All-Layer Connector for Robust Semantic Segmentation Across Domains
SummaryThis paper proposes a Sparse All-Layer Connector (SALC) for domain generalised semantic segmentation, which lets each layer access earlier layers of another foundation model. SALC adaptively selects a sparse subset of informative layers, and a candidate dropout regularisation encourages sparsity and more robust selection. The authors report stronger generalisation than conventional depth-aligned fusion across four model combinations and three evaluation settings.
WhyLearns sparse cross-depth layer selection beyond depth-aligned fusion, validated across four model combinations and three settings.
In briefA test for whether a network has a real central structure
SummaryThe paper proposes a network-level hypothesis test that uses the adjacency spectral radius as its statistic and estimates an empirical p-value from randomized networks under a specified null model. It evaluates the framework on simulated benchmarks using G(N,M), Chung-Lu, and two configuration-model implementations, plus 55 real networks. Results show a clear negative association between central-structure strength and the p-value, and rejection decisions depend substantially on which structural properties the null model preserves.
In briefDynamic conflict clustering with learning and repair for distributed constraint satisfaction
SummaryThis paper proposes DCC-Incomplete, a repair-based solver for Distributed Constraint Satisfaction Problems in multi-agent systems. It groups unsolved conflicts into clusters, limits communication to agents within the same conflict region, and adds memory-guided restarts to recover from stagnation. On distributed graph-colouring benchmarks with planted satisfying assignments, memory-guided restart raised the success rate from 37.50% to 61.11%, reduced median runtime from 0.565s to 0.367s, and cut median communication from 12.493 to 6.503 messages per constraint.
In briefSurgCast: Predicting Future Surgical Tool Skeletons from Robot Actions for Controllable Surgery Video
SummarySurgCast predicts future surgical instrument skeletons from current structure, robot state, and planned actions, then uses this forecast to guide video generation. It combines skeleton geometry with instrument-specific motion descriptors through two injection paths. On the SutureBot and SRTH-Porcine-Cholei datasets, both teacher and student models outperform direct-action baselines on all seven reported metrics, with teacher FVD dropping 37.6% and 67.3%.
In briefRecall What Is Missing: Building Complete Evidence from Long-Term Agent Memory
SummaryUnder a limited context budget, a long-term agent must select memories that together are sufficient for the current query, not just individually relevant. The paper proposes MemPursuit, a training-free method that uses the residual of a sparse convex query reconstruction to estimate what the current memory set still lacks, and composes nested evidence sets step by step. On the LoCoMo benchmark with a fixed budget, it improves end-to-end accuracy by 7.2 points at Top-10 and 5.5 points at Top-20, and runs median reranking 32.9 times faster than fully batched BGE-v2-m3.
WhyReframes long-term memory reading as evidence composition via sparse convex residuals, with LoCoMo gains and latency comparisons against a strong reranker.
In briefContext-aware precision scheduling for small language models under latency and accuracy budgets
SummaryThis paper studies how small language models running on devices should balance accuracy and latency as context length grows. The authors calibrate a timing model from measurements at different precisions and context lengths, and combine it with tensor sensitivities to choose mixed-precision configurations. They evaluate the resulting schedules on a laptop GPU, including a tool-use benchmark, and report where context-aware precision helps and where accuracy recovery remains limited.
In briefMERIT: Scaling Knowledge Graph Foundation Models with Transformers
SummaryThis paper introduces MERIT, the first knowledge graph foundation model (KGFM) with a transformer relation encoder, which predicts links zero-shot on unseen graphs. Every relation attends to every other relation, and motifs are used as a soft prior instead of edges. Across 56 benchmarks, MERIT matches the zero-shot MRR of state-of-the-art KGFMs while running 6x faster than TRIX and over 100x faster than FLOCK at inference.
Cosimo Gregucci, Daniel Hernandez, Antonio Vergari et al.
In briefBig-Big-SycoBench: A benchmark for detecting sycophancy in how models respond to valid and invalid arguments
SummaryThe paper introduces Big-Big-SycoBench, a benchmark that measures sycophancy in large language models during evaluative tasks without ground-truth answers. It uses 450 text artefacts and 5,400 human-reviewed arguments, with each model scoring artefacts repeatedly to establish a baseline before being given valid or invalid criticism. Score changes count only when they exceed the model's own baseline variability, separating responsiveness to sound arguments from compliance with unsupported pressure. The authors report that responsiveness and compliance vary widely across ten models.
Marthe Ballon, Arne Vanhoyweghen, Seorin Kim et al.
In briefThe Price of Precision: A Calibrated Cost Model for Mixed-Precision LLM Serving
SummaryThis paper argues that using average bit-width or memory budget to estimate speed is unreliable for mixed-precision LLM serving, since the limiting resource may be compute, weight loading, or KV cache access. The authors build a cost model that accounts for the binding resource to predict serving latency across hardware classes. They combine modeled time costs with measured per-tensor accuracy costs and use dynamic programming to choose configurations under latency or accuracy budgets, comparing against byte-aware and sensitivity-only allocation.
In brief1Step-WAM: A Faster World Action Model Using One-Step Action Generation
SummaryWorld Action Models unify visual world modeling and action generation, but iterative action denoising and sequential processing slow real-world deployment. The authors observe that action-denoising trajectories are approximately straight, so the first-step velocity estimates the final action direction well. Their 1Step-WAM framework uses a lightweight calibrator and bandwidth-constrained action decoding to generate actions in one forward pass that runs in parallel with visual prefilling. Evaluations on RoboTwin and LIBERO report that it largely preserves multi-step task success with about 13.4M trainable parameters and speeds up inference by 7.44x, 5.98x, and 24.84x on FastWAM, ImageWAM, and Kairos.
WhyCompresses multi-step action denoising into one step via a calibrator, validated on RoboTwin, LIBERO, and real-robot experiments with measured speedups.
In briefAdaptive role assignment and controlled feature reconstruction for federated image semantic communication
SummaryThis paper proposes ARC-FedSC, a framework that coordinates user role assignment and server-side controlled feature reconstruction under a fixed communication budget in radio access networks. Its Role-Bandit component selects roles using historical contributions and current uplink conditions, while Temporal Consistency Alignment regulates reconstruction based on recent model aggregation directions. Experiments on CIFAR-10 with time-varying channels report that it outperforms baselines in later training stages and reduces harmful reconstruction events from 0.245 to 0.093.
In briefFrom Stages to States: Unified Iterative Reasoning for Multimodal LLMs
SummaryThis paper proposes ReState, a framework that reframes multimodal chain-of-thought reasoning in multimodal large language models as a state-driven process rather than separately optimized stages. It has two components, Unified Reasoning State Modeling and Closed-Loop Reasoning-State Learning, which handle reasoning construction and correction within a shared state space. The authors also build VersaCoT-10K, a chain-of-thought dataset covering eight visual reasoning domains. They report strong performance across several MLLM backbones and public reasoning benchmarks.
In briefScreening collapse when synthetic data is recursively generated
SummaryThis paper studies how variable screening, the stage before model fitting, degrades when synthetic data is recursively generated and each generation uses pseudo-labels from the previous one. The authors call the resulting loss of variables that clean-label screening would keep "screening collapse." In a Gaussian linear model they give finite-sample bounds, a late-generation information-loss limit, and a minimax impossibility result. They also compare fixed-fraction real-label retention with data accumulation as mitigations, and gene-expression experiments support the theory.
In briefAdaptive AI for Reliable Fetal Brain Measurements on MRI
SummaryThis study presents BioTTA, a framework that lets pretrained AI models adapt online to individual unlabeled fetal brain MRI volumes, reducing accuracy loss from differing image sources. The retrospective study used 394 three-dimensional fetal brain MRI volumes from 20 to 39 gestational weeks, with 11 measurements annotated by radiologists. Across three target datasets spanning different sequences, centers, and countries, BioTTA outperformed comparison methods and matched models finetuned with target labels.
In briefProML: a framework that separates label fitting from structural regularization in multi-label learning
SummaryThis paper proposes ProML, a progressive multi-label learning framework that uses an auxiliary predictive variable to decouple direct label fitting from predictor learning. The authors theoretically analyze its optimization geometry, showing it enlarges the stable learning rate range under quadratic regularization. They instantiate ProML for multi-label feature selection and report experiments on 12 benchmark datasets that support the analysis and show competitive performance.
In briefJoint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation (JCAR)
SummaryThe paper proposes JCAR, a two-stage framework for unsupervised domain adaptation that learns a shared representation from unlabeled source and target samples by combining contrastive learning with adversarial domain alignment. The representation is then frozen and a predictor is fit using labeled source data. The authors derive a non-asymptotic excess-risk bound and report transfer gains on Office-31, Office-Home, and DomainNet, especially when representations are learned from scratch.
Yuling Jiao, Lican Kang, Jerry Zhijian Yang et al.
In briefLesion-aware attention for mammogram segmentation using mixed supervision
SummaryThis paper studies open-vocabulary lesion segmentation in mammography, where pixel-level masks are costly and only some lesion types have them. The authors combine segmentation data for base categories with image-level labels for broader categories in a two-branch framework. A Lesion-wise Attentive Classifier generates lesion-aware attention maps to pool foreground features. Experiments on multiple mammography datasets report consistent gains over baselines on base and novel categories and a narrowed gap to fully supervised models.
In briefSqueeze-Aware-YOLO: Lightweight Real-Time Traffic Sign Detection with Channel Recalibration and Multi-Scale Fusion
SummaryThis paper proposes Squeeze-Aware-YOLO, a lightweight framework for detecting traffic signs in real time. It combines a multi-branch bottleneck with Squeeze-and-Excitation channel attention and adds multi-scale feature fusion modules. The authors report mAP@0.5 scores of 74.7%, 79.4%, and 84.0% on three traffic sign datasets, outperforming recent YOLO-based detectors while keeping inference efficient.
In briefCondMSP: Condition-aware, resource-aware planning for inorganic materials synthesis
SummaryThis paper proposes CondMSP, which extends materials synthesis planning from predicting recipes to planning that accounts for execution. It predicts temperature, time, and atmosphere, using them both as process context for generating and retrieving operation sequences and as constraints on which thermal stages can share a furnace run. Under limited furnace budgets, it jointly selects routes and allocates thermal stages via constrained optimization, reporting higher recovery of reference recipes.
In briefFireNav: Routing reconstruction evidence to improve navigation in indoor fires
SummaryIndoor fire smoke and haze degrade the visual input that vision-language navigation robots rely on. The paper proposes FireNav, which uses a lightweight router to decide whether image-reconstruction features should be added as auxiliary evidence, trained according to how much they help predict navigation actions. The authors also introduce FireVLN-Bench, a paired indoor fire navigation benchmark. Experiments report that FireNav outperforms state-of-the-art baselines in navigation success and path efficiency.
In briefAttention-guided graph neural network for recovering boundary data and rebuilding interior physical fields
SummaryThis paper proposes a deep learning framework that uses an attention-guided graph neural network surrogate to recover missing boundary information. A variational autoencoder learns the nonlinear mapping from known to missing boundary quantities, and the interior field is reconstructed with the boundary element method integral formulation. Numerical tests on two-dimensional Laplace and Poisson equations report good reconstruction accuracy, local detail preservation, and model stability.
In briefThe gradient top-k distillation never computes: a proof that renormalization hides tail mass
SummaryMany deployed distillation pipelines show students only the teacher's top-k probabilities, renormalized before minimizing divergence. The paper proves that renormalization removes visibility of tail mass but not the freedom to move it, so the gradient along tail-controlling directions is exactly zero. On seven open-weight teachers, the retained set determines the teacher's distribution only when k equals d+1, where d is the hidden dimension, and commonly cached sizes fall below this.
WhyProves renormalization hides tail mass, shows zero gradient along tail directions, and validates an exact recovery threshold k=d+1 across seven teachers.
In briefInverse garment design: recovering simulation-ready 3D outfits from one image using visual agents
SummaryThis paper presents DressAgent, a hierarchical agent system that reconstructs simulation-ready 3D garments with multiple layers from a single image. It first plans layer order and estimates coarse sewing patterns, then refines them in a closed loop of editing, simulating, and judging against the reference. The authors also release a multi-layer sewing pattern dataset with over 15K outfits.
In briefEllipsoid Fitting from Tangents: detects and fits multiple ellipsoids using tangent lines and contact points, working in all dimensions and beating prior methods in 2D.
SummaryThis paper presents an algorithm that detects and fits multiple ellipsoids in 2D, 3D and higher-dimensional data. It represents ellipsoids using tangents and their contact points rather than points alone, so each tangent gives d independent equations versus one for a point, reducing the samples and iterations needed. The problem is posed as linear equation sets, with ellipsoid parameters recovered robustly by clustering in parameter space. The authors report the method works in all dimensions and outperforms the state of the art in 2D, and note it applies more broadly to robust solution of linear equations.
In briefCalibration data for pruning experts in Mixture-of-Experts models
SummaryThis paper studies how calibration data affects post-training expert pruning in Mixture-of-Experts (MoE) models, using gpt-oss-20b as the example. It examines how calibration set size, language, and data distribution influence layer-wise expert pruning. The authors report that scaling calibration data beyond small sizes gives little benefit for domain-general pruning, that English-only calibration transfers well across seven languages, and that diverse data resembling the pretraining distribution performs comparably to or better than test-distribution data.
Luis Frentzen Salim, Chia-Chun Wu, Tran Van Nhiem et al.
In briefDiffusion model that aligns its internal representations to generate video background music
SummaryThis paper proposes U-SRA, a diffusion U-Net that during training aligns weaker early-layer representations with better later-layer representations to improve internal representation quality. The method is applied to generating symbolic background music for videos, using cross-attention to synchronize music with video segments. The authors report that on a benchmark video-music dataset, U-SRA surpasses state-of-the-art methods on key metrics, including video-music alignment and overall music quality.
In briefStudies how pruning and recovery training interact when compressing large language models
SummaryThis paper examines how structural pruning and the subsequent recovery stage affect each other when compressing large language models. The authors vary the stage order, the data used in each stage, and how their training objectives align. They find that instruction-following fails first under pruning, that in-domain data is the strongest pruning signal, and that instruction-style data recovers the most, while stage order matters only slightly. The paper also analyzes how importance rankings shift before and after pruning.
Luis Frentzen Salim, Komang Rinartha, Chia-Chun Wu et al.
In briefLatent-space reweighting for prediction when input distributions shift
SummaryThis paper proposes a transfer learning method for covariate shift that combines sample reweighting with representation learning. Instead of modeling in the high-dimensional input space, it estimates density ratios of learned low-dimensional latent features for weighted prediction on the target domain. The authors derive explicit target-domain excess-risk rates and report experiments on simulated and real datasets showing gains, along with sensitivity to initialization and domain choice.
In briefTruncation samplers are nested, so comparing decoding rules becomes a measurement problem
SummaryThe paper observes that almost all deployed truncation samplers keep a prefix of the probability-sorted vocabulary, so any two such rules are nested at every decoding step. This yields a per-step distance, a coupling under which both samplers produce identical text until they first disagree, and a separation certificate bounding their difference. Experiments on six open-weight models and four math and science benchmarks find the certificate is usually at least 0.998, while a paired estimator reduces variance, so choosing a sampler is largely a measurement problem.
In briefGenesis: Visual augmentation primitives for multimodal chain-of-thought reasoning
SummaryThis paper finds that chain-of-thought reasoning in multimodal language models stays mostly in language and rarely ties each step to visual evidence, which it attributes to missing training data. The authors propose Visual Augmentation Primitives (VAP), a pipeline that uses vision generative models to synthesize photorealistic training samples with matching reasoning text, and add rubric-based reinforcement learning that scores intermediate steps for validity, completeness, and consistency. A 4B model built on this framework is reported to surpass 8B baselines, with performance improving as the model scales to 32B.
In briefUNISON: A Unified Framework for Learning from Scored Bags
SummaryThis paper introduces UNISON, a framework that unifies multiple instance learning and recommendation systems as learning from collections of scored bags of (item, score) pairs. It learns universal item representations, then bag-specific lightweight personalized parameters, and infers bag-level characteristics from them. The authors report results matching or exceeding existing methods on books, movies, T cell receptor classification, and microbiome data, and a synthetic experiment showing the bag-specific parameters recover the latent factors behind the scores.
Ofek Kirshenboim, Or Shkuri, Shani Finkelshtein et al.
In briefThompson Sampling for Learning in Zero-Sum Games: a learning method that samples a possible game and plays its equilibrium
SummaryThis paper studies Thompson sampling, an algorithm that samples from a posterior belief, in an unknown two-player zero-sum matrix game with Gaussian bandit feedback. Under centralized self-play, it samples one payoff matrix from the posterior and plays a Nash equilibrium of that sampled game. The authors show the expected cumulative duality gap is O(√T), improving to O(log T) for pure equilibria, and provide matching lower bounds and experiments.
In briefA plausible future isn't necessarily a causal one: can video models generate how the world unfolds?
SummaryThis paper introduces CausalCascade, a benchmark built on 50 event-rich Blender scenes with structured event graphs and 2,832 counterfactual variants. It tests whether image-to-video models can faithfully realize a full causal chain from an initiating event to a final outcome. Across nine evaluated models, generated motion is more reliably localized in space than correctly evolved over time, and responsiveness to physical interventions is weak.
In briefVision-Predictor-Action: A faster robot control method using neuro-symbolic JEPA predictors
SummaryThis paper proposes Vision-Predictor-Action (VPA), which replaces the slow token-by-token action decoders in vision-language-action robot models with a lightweight latent control loop. It combines a neuro-symbolic skill selector, a JEPA latent predictor, and a flow-matching solver to generate a whole action chunk in a few steps. On the LIBERO-10 suite, the trainable model has under 20M parameters, and a two-camera policy succeeded in 40.0% of 200 evaluation episodes. The authors describe these as preliminary results from a single training run.
In briefELECT: Fixing Character Identity Before Rendering for Scalable Story Visualization
SummaryThis paper introduces ELECT, a training-free method for story visualization without reference images. It settles recurring character identity across frames before final rendering, using language-guided scene-entity mapping and stationary anchors chosen from the model's own drafts, so per-frame rendering cost does not grow with story length. On the Story50 long-story benchmark, ELECT achieves higher character consistency than ten-round refinement at about one seventh of the generation cost. The authors also report results at 1000 scenes.
Mehrajul Abadin Miraj, Md. Mahmudul Hasan, Meherun Farzana et al.
In briefAn interpretable, linguistics-based analysis of which movie screenplay scenes are salient
SummaryThis paper studies which linguistic factors make scenes in movie screenplays salient, using a controlled study on the MENSA corpus. It proposes LGSM, a classifier that combines a frozen BERT encoder with explicit linguistic features through adaptive gating. The authors find that scene-level Uniform Information Density summaries carry no independent signal once length is controlled, but at the token level salient scenes suppress the rarest, hardest-to-predict tokens. Salient scenes also show sustained central-character presence, more third-person usage, simpler syntax, and more abstract vocabulary about motivation and conflict.
In briefClient-specific memory windows for federated learning under drift
SummaryThe paper studies federated learning with continuously arriving data, where each client's distribution drifts over time and the server tracks the minimizer of the weighted average loss. It gives each client its own memory window sized to its noise, arrival rate, and drift, refreshed by a restarted accelerated method from a common point. The derived tracking-error bound is never worse than any shared window and is strictly better when noise and drift are unevenly spread, and experiments attribute the gain to the memory rather than the solver.
Yuriy Dorn, Dmitrii Vishovan, Kirill Novoselov et al.
In briefApertus 1.5: Open Multimodal Reasoning Models Built on Open Text Models
SummaryThis paper presents Apertus 1.5, which extends the fully open Apertus family from text to multimodal reasoning. Continued pretraining adds image and audio understanding and expands the context window to 262,144 tokens, with thinking enabled as a chat template flag. The authors report improvements over predecessors on nearly every benchmark and release model weights, data recipes, and evaluation suites under permissive licenses.
SummaryThe paper proposes sensitivity fidelity, an evaluation criterion that checks how models respond to input perturbations against a trusted reference or physical knowledge, in addition to forecast error. In an idealized Rayleigh-Bénard convection benchmark, four surrogate models had low forecast errors but very different sensitivities, and comparison with a matched adjoint distinguished them more sharply than accuracy did. Applied to El Niño-Southern Oscillation prediction, where no adjoint exists, the learned sensitivities were robust across events and consistent with established precursors.
In briefUMAS3D: Unified Mask-Free 3D Editing That Preserves Unedited Source Regions
SummaryThe paper presents UMAS3D, a unified framework for editing existing 3D assets that needs no training, no inversion, and no external 3D masks. It converts text, bounding boxes, and reference images into shared target-view guidance so that adding, deleting, modifying, appearance, and pose edits use the same generation process. The method estimates the edit region online and constrains the remaining region to keep source geometry and appearance. Experiments reportedly show it supports these edit types while preserving unedited content.
In briefAgent-Canary-Mas: Measuring How Protocol Attacks Spread in Long Multi-Agent Tasks
SummaryThis paper presents Agent-Canary-Mas, an evaluation system that tracks how risks from agent-to-agent protocols carry through long, multi-step tasks involving several AI agents. It covers five risk classes across A2A, ANP, and BCS protocols, using 975 matched malicious and benign task pairs in five business domains. The authors find that local attack success is often at or above 90%, but conversion into real consequences ranges from 7.1% to 100.0%. They also test a protocol-context guard, which on Finance tasks lowered attack success and downstream consequences while reducing task completion.
In briefA Step in the Right Place: Optimizing Timestep Placement in Continuous Diffusion Language Models
SummaryThis paper examines how the placement of denoising timesteps affects continuous diffusion language models, even when the number of steps is fixed. The authors show that changing only timestep placement alters the final latent representations and decoded token distributions. They propose StepTune, which searches for better placements on a small calibration set using a language-aware objective, then reuses the placement across inputs. Reported results include better perplexity and task metrics across sampling budgets, and 16-step sampling outperforming a 64-step baseline on XSum and WMT14 with about 3.8x lower latency.
In briefGNFormer: Gaussian Neural Field Transformer for camera-based occupancy prediction
SummaryThis paper proposes Gaussian Neural Field (GNF), a 3D representation where Gaussian primitives and a neural network jointly map voxel positions to semantic occupancy. Building on it, the authors present GNFormer, which predicts GNFs from camera input. On the Occ-ScanNet, SSCBench-KITTI-360, and nuScenes-SurroundOcc benchmarks, it achieves the highest IoU and mIoU and the fastest inference among compared single-frame methods. The paper reports representing indoor and outdoor scenes with only 256 and 1024 Gaussians.
In briefTighter approximation bounds for weighted maximin share fairness in allocating chores
SummaryThis paper studies how to fairly divide indivisible chores among agents with different weights and nonnegative additive costs. An allocation is α-WMMS fair if each agent's cost is at most α times her weighted maximin share. The authors prove that every instance admits a 4.124234528-WMMS fair allocation and give a polynomial-time algorithm for rational inputs that finds a 4.12424-WMMS fair allocation. They also construct instances where the optimal approximation factor approaches e, narrowing the previously known gap from [2,12] to [e,4.12424].
Trung Thanh Nguyen, Khaled Elbassioni, J\"{o}rg Rothe
In briefOceanFlow: Coherent Flow Matching for Ocean Forecasting Under Uncertainty
SummaryThis paper presents OceanFlow, a flow matching framework for probabilistic global ocean modeling that produces ensembles of future ocean states. It first sets a stable mean anchor, then learns a state-conditioned pathwise law that reshapes residual uncertainty around that anchor. The authors also introduce a Factorized Dynamics Operator (FDO) to parameterize the anchor and residual-flow velocity. Experiments report gains over competitive baselines in ensemble-mean accuracy, probabilistic reliability, and physical-consistency diagnostics.
In briefSemantically Steerable Image Tokenization: Keeping What an Image Depicts When Compressing
SummaryThe paper introduces GIST, an adaptive image tokenizer that allocates tokens using jointly controlled pixel-level and semantic reconstruction targets. Semantic error is measured in the feature space of a frozen DINOv3 encoder, allowing pixel detail to be discarded while semantic content is preserved. In a controlled comparison on ImageNet-100, GIST outperforms the pixel-controlled KARL at every token count, with up to 30% lower LPIPS and 37% lower DreamSim.
Danilo de Goede, Mohammad Mahdi Derakhshani, Yuki M. Asano et al.
In briefSARATHI: A method for structured LLM pruning that separates data-free neuron probing from adaptive weight reconstruction.
SummaryThis paper presents SARATHI, a framework for structured pruning of large language models to reduce inference cost. It separates neuron selection from weight reconstruction: a data-free NMF residual subspace probe scores neurons using weight geometry, and adaptive OBS then reconstructs pruned weights using a localized calibration batch. A bias shift compensation step corrects activation drift in post-LayerNorm architectures. Experiments on six models from two architecture families at 15–40% sparsity report competitive perplexity and zero-shot accuracy matching or exceeding existing structured pruning baselines.
Sankar Behera, Dhruv Singh, Satyadev Ahlawat et al.
In briefREPRO: A Reflective Evolutionary Method for Automated LLM Red-Teaming
SummaryThe paper presents REPRO, a pure black-box framework that finds adversarial prompts able to bypass the safety training of language models, using evolutionary optimization driven by a compressed 7B open-weight model on a single consumer GPU. It adds semantic crossover and Thompson Sampling-based strategy selection informed by failure analysis. The authors report 100% single-turn attack success rate across 36 target models, but note that the judge used during optimization also produces the reported figure and baseline numbers come from original papers with different protocols, so it is a model-coverage catalog rather than a head-to-head ranking.
In briefStructured scripts and a dataset for consistent multi-shot audio-visual video generation
SummaryThis paper addresses keeping identities consistent, tracking state changes, and coordinating sounds across shots in multi-shot audio-visual video generation. It proposes Structured Stateful Relational Scripts (SSRS), which separate persistent entity and voice identities from shot-specific visual and auditory states and explicitly encode sound-source bindings and temporal relations. The authors also introduce MSAV-Script-100K, a dataset of curated multi-shot audio-visual videos with detailed SSRS annotations, and report improvements in cross-shot consistency, state-change adherence, and audio-visual alignment.
In briefGeometric priors for neural Calabi–Yau metrics on the quintic: a fixed-ψ ablation study
SummaryThis paper studies how geometric priors affect neural approximations of the complex Monge–Ampère equation on the quintic Calabi–Yau threefold. It compares four single-patch models: a coordinate SIREN potential, a direct-metric Cholesky baseline, an algebraic section ansatz, and a section-plus-neural hybrid. Section-based models reach validation σ of about 4.14–4.16%, while SIREN and Cholesky remain at about 41.29% and 44.27%.
In briefCacheLink: Functional KV Transport for Zero-Prefill Cross-Model Context Handoff
SummaryThis paper studies how to move a long-context KV cache from one language model to another without re-prefilling the context. It proposes CacheLink, which learns a coupled K/V correction by directly optimizing the receiver model's behavior rather than only reconstruction error. The authors report that reconstruction fidelity alone is a poor proxy for receiver accuracy, and that at 32K context the method cuts handoff latency by up to 7.27 times versus re-prefill.
In briefRethinking Data Portability for Human-Centric, AI-Driven Data Ecosystems
SummaryThis paper critically examines the current US regulatory framework and debate on data portability, using recent European Union developments as an illustration. The author argues that existing legal conceptions of portability mainly serve the interests of service providers and data controllers rather than individual users. It calls for a shift toward a human-centric approach in which individuals control their own data, and considers how EU technical and market trends could support such a change.
Mark Fenwick, Michael Fertik, Paulius Jurcys et al.
In briefLarge Decision Models: Using LLMs as single-pass decision makers with explicit label probabilities
SummaryThe paper proposes Large Decision Models (LDMs), which adapt language models to return explicit probabilities over run-time-specified labels in one prefill forward pass, avoiding sequential decoding. It uses Diverse Decision Supervision and Self-Posed Reasoning Distillation to improve performance. On DecisionVista, the method improves accuracy over a step-matched baseline by 1.8 and 1.5 points on two base models and gives a 182-fold latency speedup on Gemma-4-31B. A confidence router is also proposed to send uncertain decisions to reasoning, which the authors report recovers most of reasoning's accuracy advantage while generating a fraction of its tokens.
In briefThe Unerasable Record, Forgiveness, and the Future of Criminal Justice
SummaryThe paper argues that digital systems can keep records of wrongdoing indefinitely, and AI can let those records keep shaping a person's future, a condition the authors call the "eternal data present" that threatens forgiveness. Drawing on Hegel's "hard heart" and Arendt's irreversibility of action, it proposes a "right to remain unfinished," including limits on the weight of older records, re-evaluation after personal change, reassessment of inferences affecting people's lives, and time limits on using records in specified decisions.
Mark Fenwick, Paulius Jurcys, Sebastian Porsdam Mann
In briefRevisited SVM: A Unified Joint Optimization Framework for Scalable Adaptive Learning
SummaryThis paper proposes a "Revisited SVM" framework that jointly optimizes dual Lagrange multipliers and kernel hyperparameters instead of relying on grid search. It combines a hybrid block coordinate ascent algorithm with Nyström low-rank approximation, reducing computational complexity to O(Nm). The authors report experiments in bioinformatics, finance, cybersecurity, and computer vision, including 96.25% accuracy on the NSL-KDD dataset.
Nassou Gongon Benjamin, Noulapeu Ngaffo, Patrick Njionou Sadjang
In briefMeasuring how much people share the meaning of a concept, using sentence embeddings
SummaryThis paper proposes a framework that uses sentence embeddings to measure how differently individuals understand the same concept. Short definitions from 125 participants were compared pairwise to compute sharedness and idiosyncrasy scores. The authors report that the approach is validated and that the results replicate across three different encoders.
In briefIdentifying Signal Processing Software Using Machine Learning on Over-the-Air Radio Signals
SummaryThis paper examines whether transmitters can be identified by the digital signal processing (DSP) software that generates their signals, rather than by hardware imperfections. Five software implementations producing a 2-FSK RTTY waveform are classified with convolutional neural networks, tested on over-the-air recordings from a shortwave transceiver and software-defined radio. The best model, combining IQ and power spectrum inputs, reaches up to 90% accuracy, and the study also reports effects of signal SNR and observation bandwidth.
In briefSkills Remain, but Access Fails: A Factorized Memory for Continual Skill Learning
SummaryThis paper studies how language-model agents lose previously internalized skills when they continually learn new ones. The authors find forgetting occurs in two stages: first the model can no longer invoke an old skill from the raw request, and later even the correct skill name no longer yields good execution. They propose Factorized Skill Memory (FSM), which separates a plastic Address Memory from a stable Execution Memory of frozen LoRA executors. On a five-stage benchmark built from 100 TaskBench skills, FSM reports 65.95% mean exact match.
In briefAlphaPilot: RL-Trained Multi-Turn LLM Agents for Adaptive Alpha Mining
SummaryAlphaPilot is a reinforcement-learning framework that trains multi-turn, tool-using LLM agents to mine interpretable formulaic alpha factors for ranking assets by future returns. The agent combines market descriptions, historical data, and evaluator feedback, updating a retained factor pool across turns and searching day by day at inference. In out-of-sample rolling backtests on CSI300 and CSI500, the paper reports mean total returns of 42.93% and 54.19% and mean Sharpe ratios of 2.46 and 2.86 over three independent runs.
In briefSkillFusion: A Method for Reliably Combining Multiple Skills in LLM Agents
SummarySeparately written skills can conflict, compete, or repeat instructions when an LLM agent loads them together. SkillFusion is an offline compiler that reconciles these interactions into one reusable merged skill, making composition decisions explicit through typed, scoped constraints linked to their source directives. It handles hard conflicts, soft conflicts, and redundant instructions differently. The paper reports the highest mean Success among non-oracle methods across three agent benchmarks, two transfer benchmarks, and three executor models.
In briefRepresentation theorems for actual powers in two-agent general concurrent game frames
SummaryThis paper generalizes the classic link between logic and games, showing that with two agents, eight classes of general concurrent game frames are represented by eight corresponding classes of neighborhood frames in terms of actual power. The eight classes arise from combinations of three assumptions: seriality, agent independence, and determinism. The result generalizes an earlier theorem by van Benthem, Bezhanishvili, and Enqvist (2019).
Zixuan Chen (Institute for Logic, Language, Computation et al.
In briefA semantic saliency method using multimodal LLM scene descriptions
SummaryThe paper introduces AUTO-SUM, a fully automated method that measures each object's contribution to scene understanding. It segments and removes objects one at a time, has multimodal large language models describe the original and modified scenes, and compares the descriptions using text embeddings. The authors report that AUTO-SUM predicts the most important object in human scene understanding maps and the object most fixated by observers, outperforming low-level saliency models and the Meaning Map, and performing comparably to DeepGaze.
Ziqi Wen, Shravan Murlidaran, Jonathan Skaza et al.
In briefMemeLogos: A Knowledge-Driven Retrieval-Augmented Framework for Detecting Harmful Memes
SummaryThis paper addresses the difficulty of detecting harmful memes, whose harmfulness is often conveyed implicitly through visual, textual, and contextual cues. The authors propose MemeLogos, which abstracts recurring harmfulness patterns and reasoning structures into reusable knowledge units organized in a star-shaped knowledge graph. The framework combines structured knowledge retrieval with a multi-stage reasoning process. According to the authors, it outperforms existing methods while producing more coherent and interpretable reasoning.
In briefScalable robust tree search for planning under model uncertainty, with guaranteed bounds
SummaryThis paper addresses planning in Markov decision processes where the transition model may be inaccurate. The authors propose CAR-MCTS, a Monte Carlo tree search planner that computes only on a sampled tree while deterministically bounding the effect of unsampled trajectories. The bounds are available at any time, tighten as more of the tree is sampled, and can certify the optimal root action or the robust value of a policy. The authors claim this is the first scalable robust online planner with anytime deterministic guarantees.
In briefWhen Messages Go Wrong: Communication Faults in LLM Multi-Agent Systems
SummaryThis paper studies how communication faults such as message loss, delay, and context expiration affect LLM-based multi-agent systems, where agents exchange natural language messages. The authors build a cross-layer taxonomy and a fault-injection and tracing pipeline, then evaluate over 29,000 executions across 100 tasks and three workflow topologies. They find that many failures occur without an immediate failure signal, and that workflow topology strongly affects task success under faults.
In briefConceptizer: Next-Token Prediction Is Not Enough
SummaryThis paper introduces Conceptizer, a training framework that adds a reward for meaning to standard next-token prediction. It uses a fixed vocabulary of about 9K concepts, built from contextual synonyms, and maps next-token distributions to concept distributions so that alternative wordings of the same meaning share the reward. Tested on three model families at 1B to 8B parameters, it achieves the lowest held-out perplexity at concept-target positions and improves downstream knowledge and sentiment tasks, while remaining comparable on commonsense.
In briefInterval Length of Split Conformal Prediction under Huber Contamination
SummaryThis paper studies how Huber contamination in training and calibration data changes split conformal prediction interval length compared with the clean population oracle. It derives a finite-sample upper bound that holds uniformly over contaminating distributions, and shows a matching minimax lower bound on a regular asymmetric location submodel. For calibration contamination, it gives an exact worst-case characterization of the population interval-length deviation, and the upper bounds are extended to conformalized quantile regression and adaptive residual scores.
In briefPMB-NN: A Physiology-Constrained Neural Network for Predicting Blood Glucose in Type 1 Diabetes
SummaryThis paper presents PMB-NN, a method that pairs a temporal convolutional network with a personalized four-parameter physiological model to forecast glucose levels in type 1 diabetes. It uses 6 to 8 weeks of continuous glucose, insulin pump, and meal data from 25 participants, predicting 30, 60, and 120 minutes ahead. Without basal insulin records, PMB-NN had the lowest median error at 30 and 60 minutes and outperformed the physiological model alone in all six settings.
In briefBridgeCodec: Mamba-enhanced neural audio codec with Schrödinger Bridge for low-bitrate audio
SummaryNeural audio codecs usually pair an encoder and decoder that are trained together, which limits interoperability. This paper proposes BridgeCodec, which translates latent representations between mismatched frozen endpoints using a Schrödinger Bridge optimal transport formulation and a Mamba-enhanced U-Net for long-range speech dependencies. It is tested by mapping a lightweight 8 kHz encoder to a 48 kHz decoder at 1 kbps. The reported results show strong wideband reconstruction and perceptual quality.
In briefEmoTrace-Bench: A Diagnostic Benchmark for Emotional Expression in Text-to-Video Generation
SummaryThis paper introduces EmoTrace-Bench, a fine-grained benchmark that controls emotion intensity and selected affective cues in prompts to diagnose how text-to-video models express emotions. It uses 448 prompts and 12 models to generate 5,376 videos, evaluated with an automated evaluator validated against human annotations. The results show that prompts explicitly describing a character's emotional reaction are followed more reliably than those describing only events or scenes, while adding more cues or explicit cinematic language does not consistently help.
In briefHidden in Plain Sight: Teaching AI to Catch Disguised Harmful Content
SummaryThis paper notes that large language models can miss harmful intent that humans readily recognize, which malicious users exploit to evade moderation. The authors propose PASP, which builds textual, phonetic, and visual views of each input and trains an attacker and defender through adversarial self-play. Experiments report that PASP reduces adaptive attack success rates by up to 73.71% on ChineseHarm-Bench and 74.49% on EvoHarmBench, with minimal loss of model capability.
In briefRethinking supervision for document unwarping with a flow-free, self-consistent method
SummaryPhotos of documents taken with smartphones are often distorted, which hurts digital preservation and downstream use. The paper introduces DRNet, a deep network that rectifies document images using a coarse-to-fine design. It supervises training with undistorted document images instead of warping meshes used as ground truth in prior work. On the DocUNet Benchmark, the method sets a series of state-of-the-art results, and ablation experiments validate its design choices.
In briefUsing SPAD Noise to Improve Indoor Depth Estimation at Test Time
SummaryThis paper addresses challenges in monocular depth estimation, including scale ambiguity, edge bleeding, and weak zero-shot generalization. It proposes HSN, a lightweight framework that treats native single-photon avalanche diode (SPAD) noise as an informative prior rather than an artifact to remove. HSN combines 1D SPAD histograms with Gate Adapters to generate channel-wise gated attention that adjusts encoder features at inference time. Experiments on real-world indoor benchmarks report state-of-the-art metric depth accuracy in-domain and stronger potential for zero-shot generalization to unfamiliar environments.
In briefWorld Modeling through Spectral Alignment: Specifying What World Models Should Preserve
SummaryThe paper introduces SpecWM, a JEPA framework that uses a target similarity kernel to specify relationships between observations, training the encoder to reproduce them in its embedding geometry. Kernels range from physical state information to a fully self-supervised temporal kernel needing only observation order within trajectories. Experiments show the kernel choice strongly controls which physical quantities are recoverable, and temporal spectral alignment raises CEM planning success over LeWorldModel on OGBench Cube, OGBench Scene, and CALVIN.
Holger Molin, William Peng, Marco Bagatella et al.
In briefRobust Offline Meta-RL via Pessimistic Latent Task-aware Optimization
SummaryThis paper proposes PLATO, a method for offline context-based meta-reinforcement learning that handles tasks outside the training distribution. It assumes task latents lie on a low-dimensional manifold and perturbs inferred latents outward from the training centroid to create counterfactual tasks, then trains the policy on imagined rollouts from a decoder ensemble with a pessimism weight that down-weights steps where the ensemble disagrees. The authors report results on eight MuJoCo continuous-control benchmarks, where PLATO consistently improves out-of-support returns over strong baselines while staying competitive in-distribution.
In briefRiemannian graph tokenizer that turns graph structures into transferable tokens
SummaryThis paper proposes the Riemannian Graph Tokenizer (RGT), which converts graph structures into discrete tokens to support knowledge transfer across domains. It uses Riemannian geometry to define a vocabulary of trees, cycles and sequences, and a quantizer with a Riemannian straight-through estimator to tokenize graphs from multiple domains. A geometry-aligned decoder maps the tokens into a unified tangent space. The authors report that experiments on diverse datasets show significantly improved structural knowledge transferability.
In briefSpecFit: Speculative decoding that accounts for expert-transfer costs in offloaded MoE models
SummaryMixture-of-experts models offloaded to host memory suffer slow expert transfers during decoding, which can make speculative decoding slower than standard decoding. SpecFit selects the draft subtree that maximizes predicted throughput rather than acceptance length, using draft path probabilities and the union of experts activated. On Qwen3-30B-A3B, it decodes 21.6% faster than EAGLE-3 on CNN/DailyMail and 7.0% faster on a GSM8K subset, under the same expert cache.
In briefSensitivity or Criterion? Measuring and Adjusting When Language Models Act
SummaryThe paper introduces FlipAct, a diagnostic that tests whether language models should use tools in physical states, using rule-based answers and graded evidence scores. Across 44 open models, sensitivity grows with size within a family, while the decision criterion is not determined by size. Activation steering shifts the criterion, and distilling from a steered copy moves that shift into the weights without a significant sensitivity change, transferring to six external benchmarks.
WhySeparates sensitivity from criterion via signal-detection theory using flip-state 2x2 quads; tested on 44 models and several external benchmarks, giving fairly solid evidence for when-to-act behavior.
In briefOmniRSI: A Unified Latent Diffusion Model for Multimodal Remote Sensing Image Generation
SummaryThis paper proposes OmniRSI, a unified model for generating multimodal remote sensing images that recovers complete observations from heterogeneous, partially observed data. It combines a remote sensing-specific variational autoencoder, a Stable Diffusion 3.5 model with a ControlNet branch, and a training-free cyclic latent alignment mechanism. The authors also introduce RSI-MM, a dataset of 11.5M multi-resolution images, and report superior results in multimodal reconstruction, generation, and translation.
In briefFrom Candidate Tokens to Final Outcomes: Decoding Relative Long-Horizon Candidate Value from Reasoning Model Representations
SummaryThis paper asks whether reasoning models' internal representations encode how much different next-token candidates at the same prefix affect the chance of a correct final answer. The authors build the RCTV dataset pairing math reasoning prefixes with candidate tokens and empirical long-horizon values. Sparse linear probes show this relative value is linearly decodable at multiple layers and can outperform probability- and embedding-based baselines.
In briefCoverSynth: Coverage-guided generation of test sets for evaluating retrieval systems
SummaryEnterprise agents need retrieval evaluation sets that reveal domain-specific failures, but neither human-written sets nor existing synthetic methods reliably provide adequate coverage, reliability, efficiency, and discriminability. The paper introduces an assessment framework for these criteria and CoverSynth, a generator that builds evaluation sets from unannotated corpora while using deterministic and model-based checks to ensure reliable ground truth. Across five public enterprise-style corpora, the method yields varied difficulty and exposes failure modes such as handling underspecified or ambiguous requests. Its generated sets also show more consistent differences between agent configurations than human-made sets.
Anahita Sharma, Apaar Shanker, Rishav Chakravarti et al.
In briefHow state representation affects long-horizon agent learning and transfer: an empirical study
SummaryThe paper studies deterministic tool workflows in which models learn from either compact or full views of task state. A 1.7B Qwen model trained on compact views completed all held-out tasks in that view but none in the full view, while full-view training results varied across seeds. The authors conclude that representation-sensitive transfer is a bottleneck in these controlled workflows, though the observation changes were not isolated to content reduction alone.
In briefA Decade of PhotoMappers: Locating Crowdsourced Disaster Photos with AI Vision Models
SummaryThis paper studies more than ten years of PhotoMappers, a crowdsourced disaster-mapping project, to examine how the geography of disaster information has changed over time. It combines street view and satellite imagery for cross-view geolocalization and evaluates vision-language models on 30 geographic reasoning indicators. Results show uneven performance across years and models, and the authors argue that GeoAI is most valuable when it complements, rather than replaces, human geographic interpretation.
In briefUsing vision-language models as success detectors for reward design in from-scratch embodied reinforcement learning (VLMeg)
SummaryThe paper argues that training vision-language models (VLMs) as precise reward regressors needs many annotations and works poorly for from-scratch reinforcement learning. It proposes VLMeg, which uses VLMs as success detectors by first gating episodes via similarity to a generated success trajectory, then sending the key observation to the VLM for episodic reward, with added exploration incentives. Experiments in MetaWorld on 12 reward-free tasks and 3 real-world tasks report state-of-the-art from-scratch RL results.
In briefGA-SVD: a generation-aware low-rank method for compressing large language models
SummaryLarge language models are costly to run because of their size. This paper proposes GA-SVD, a one-shot low-rank compression method aimed at generation tasks, using gradients from on-policy knowledge distillation to build whitening transforms and an evolutionary search to allocate ranks across layers. The authors report that it outperforms existing SVD-based baselines on Mistral-7B, Qwen3 and Llama-3.3 models across multiple-choice and generative benchmarks. They also state it scales to models with up to 70B parameters.
In briefHow false beliefs spread, persist, and resist correction in AI agent social networks
SummaryThis paper uses the Mandela effect as a lens to study how false claims spread among communicating AI agents in a simulated social network. Using 540 fabricated claims, the authors measure dissemination, persistence after fabricators are withdrawn, and mitigation through preventive versus corrective measures. They report dissemination reach up to 96.0% and find that acceptance barely declines after withdrawal. Targeted corrective interventions outperform preventive ones but do not eliminate false beliefs.
In briefWUSPruner: Second-Moment-Optimal Transforms for Structured Pruning
SummaryThis paper introduces WUSPruner, a transform-based framework for structured pruning of large language models. It treats choosing the basis for pruning as a local linear reconstruction problem and builds a WUS transform from activation and weight statistics. Experiments on Llama and Qwen models from 8B to 72B parameters report consistent gains over structured pruning baselines, along with lower inference latency and memory usage.
In briefDC-Flow: A method for designing ring-shaped peptides that bind a target, by building the backbone first and adding amino acids afterward.
SummaryThis paper presents DC-Flow, a staged generative framework for designing cyclic peptides conditioned on a target. It first forms the target-conditioned backbone, then activates sequence and side chains and couples their flows so the two co-evolve. The authors report improved cyclization fidelity, diversity, and interface energetics, and a higher yield of confident cyclic peptide binders under a fixed final-candidate budget.
In briefVideoEvoHarness: A Framework for Self-Evolving Video Agent Harnesses from Unlabeled Videos
SummaryThis paper presents VideoEvoHarness, a method that evolves the harness, the set of tools and workflows used by video understanding agents, using only unlabeled videos and without changing model parameters. A Questioner and a Solver harness are improved jointly through execution feedback, and candidate harnesses are compared on identical validation questions with only validated improvements kept. Experiments report higher overall accuracy than the initial harness, with gains that transfer across backbone models.
In briefWhy Inner-Product Scores Approximate Contrastive Learning Targets
SummaryThis paper studies contrastive representation learning, whose population target is the pointwise mutual information between two modalities. It explains why standard inner-product scores can approximate this generally nonseparable target by linking it to a compact operator and analyzing rank-D spectral approximation. The authors also study neural network approximation rates and a calibration rate from contrastive excess risk to downstream retrieval error.
WhyLinks PMI to compact operators, giving rank-D spectral error bounds and neural consistency and rate results, a notably strong theoretical contribution.
In briefE8 lattice key quantization: a method for low-bit KV-cache compression
SummaryThis paper addresses KV-cache compression for long-context large language model inference by proposing a k2v2 scheme (2-bit keys and 2-bit values) that quantizes keys using vector quantization based on the E8 lattice. The authors report about 6x KV compression across five LLM families, roughly 20% smaller than scalar k3v2, while keeping long-context perplexity near baseline. They also state that E8 beats scalar k3v2 in 80% of model-dataset-context settings, with decode throughput matching scalar baselines.
Sasho Nedelkoski, Alexander Acker, Soeren Becker et al.
In briefDynamics of WSBM-based echo state networks analyzed with InputDSA
SummaryThe paper proposes a reservoir computing framework that replaces random networks with modular structures generated by a Weighted Stochastic Block Model (WSBM) in Echo State Networks. Input Dynamical Similarity Analysis (InputDSA) is used to compare how structural priors shape intrinsic recurrent dynamics versus external inputs. The authors find that structural diversity does not necessarily produce behavioral diversity, and that successful networks converge dynamically under environmental constraints while failing networks follow idiosyncratic trajectories.
