In briefWhich Answer Did the Seventeen-Year-Old Write? AI Models Guessing Exam Writers' Ages
SummaryThis case study tests six frontier language models on matching three handwritten exam answers to their authors' ages and genders. Most models often assign a 17-year-old's answer to an adult, and several perform near chance while reporting high confidence. The authors ran 1,081 sessions across six experimental conditions and note the study isolates a mechanism rather than estimating a population bias rate.
In briefKolibri: A European Sovereign Model on the Pareto Frontier
SummaryThis report introduces Kolibri, an English-German mixture-of-experts transformer with 78.1B total parameters and 3.46B active per token, released as open weights under Apache 2.0. It is designed for sovereign, specialised deployment, with a focus on German and regulated sectors such as public administration and aerospace. The authors state it was trained on 24T tokens, with German over 20% of the data, and that it lies on the Pareto frontier of quality and serving cost among evaluated open models in English and German.
Aleph Alpha: Moritz Althaus, Artur Baranowski, Michael Barlow et al.
In briefUncertainty-aware ambulance travel time estimates for safer AI-assisted emergency dispatch
SummaryThis paper addresses the problem that current AI-based ambulance travel time estimates give only single point predictions, without showing how reliable they are. The authors propose an uncertainty-aware method that generates context-adaptive prediction intervals. They evaluate it retrospectively on two years of ambulance trip data from Skåne, Sweden, and report stable empirical coverage across operationally important contexts.
Muhammad Adil Abid, Johan Holmgren, Fabian Lorig et al.
In briefLoop-count agreement, not state normalisation, keeps depth reusable in looped language models
SummaryThis paper studies whether looped language models, which reapply one block of layers repeatedly, can reuse depth beyond their trained loop budget. It tests whether this depends on normalising the state between loop steps or on agreement between training and inference loop counts. Experiments on released Ouro weights and small models trained from scratch suggest agreement, not normalisation itself, is the key factor.
Ivan Viakhirev, Kirill Borodin, Maxim Abramov et al.
In briefM³: A Multi-LLM System Modeling Multi-Turn Jailbreaking as a Markov Decision Process
SummaryThis paper introduces M³, an automated red-teaming framework that formalizes multi-turn jailbreaking as a finite-horizon Markov Decision Process. An attacker LLM interacts with a target LLM while using a compact internal state produced by a state LLM. According to the paper, M³ achieves higher attack success rates than prior multi-turn red-teaming methods across two benchmarks and six target models.
In briefBeyond Local Accuracy: Testing How Well PDE Foundation Models Hold Up Over Long Simulations
SummaryThis paper examines whether PDE foundation models that perform well on short predictions remain reliable when their outputs are fed back repeatedly over long horizons. The authors evaluate seven models on seven two-dimensional time-dependent PDE datasets and define an error amplification ratio comparing full-horizon to short-horizon error. They find that full-horizon error is usually much larger than short-horizon error, so short-term rankings can hide substantial amplification. The authors recommend that benchmarks report horizon-resolved absolute errors, amplification statistics, and failure rates.
In briefEach Neural Network Weight Encoded as a Path of Shared Discrete Choices
SummaryThe paper introduces TWIG, a representation that writes each neural network weight as a path of small discrete decisions such as sign, power-of-two shift, zero gate, and shared-table index, with related weights sharing upper path segments. The authors argue that pruning, quantization, factorization, and weight sharing can then be treated as edits to one object and optimized together. On five weight matrices of a large language model, they report fewer additions than GPTQ at equal layer error for 3- and 4-bit accuracy. Exact figures should be checked in the original text.
In briefText-based person anomaly search from synthetic to real images via multi-model fusion and reranking
SummaryThis paper addresses text-based person anomaly search, where people are retrieved from image galleries using descriptions of their appearance and behavior. The models are trained on synthetic image-text pairs but evaluated on real surveillance images with domain differences and similar distractors. The authors adapt Qwen3-VL-Embedding-8B to synthetic data, fuse retrieval scores from six Qwen3-VL checkpoints and a frozen SigLIP2 model, and rerank top candidates with Qwen3-VL-Reranker-8B. On the official AI City Challenge 2026 Track 4 test, the system reports 94.7815% mAP and 90.8493% R@1, with an eighth-place public leaderboard finish.
Eugene Denteh, Blessing Agyei Kyem, Andrews Danyo et al.
In briefENCL: Efficient Neuromorphic Continual Learning Using Local Learning Rules and Sub-Network Freezing
SummaryThis paper proposes ENCL, a method that makes neuromorphic continual learning more efficient. It uses a lightweight temporally and spatially local learning rule instead of costly global backpropagation, and freezes sub-networks per task to avoid expanding the network architecture. In task-incremental learning experiments on Split MNIST, Split Fashion MNIST, Split CIFAR-10, and Split CIFAR-100, it reports accuracy comparable to replay-based methods while reducing memory cost by up to 88% and energy consumption by up to 65%.
Rachmad Vidya Wicaksana Putra, Shuochen Zhou, Hassan Raza et al.
In briefCorrecting Input Bias in Gaussian Process Regression
SummaryStandard Gaussian process regression assumes inputs are known and uncertainty comes only from measurements. When a mobile sensor's pose estimates contain errors, such as from dead-reckoning, this paper proposes a bias-aware GP (Ba-GP) framework that uses the trained posterior-mean gradient to convert input bias and noise into mean and covariance corrections. First-order correction algorithms then update the model efficiently. Simulations and an indoor light-intensity mapping experiment show better predictive accuracy and uncertainty calibration than noisy-input GP baselines, with roughly an order-of-magnitude lower computational cost than full retraining.
Muzaffar Qureshi, Tochukwu E. Ogri, Trivikram Satharasi et al.
In briefBeyond On-Policy vs. Off-Policy: At What Level Should LLM Agent Distillation Follow the Student?
SummaryThis paper studies how to distill interactive LLM agents, arguing that the choice between teacher and student data should be made separately for interaction histories and current action prefixes. It proposes a hierarchical framework that controls the pairing, mixture proportions, and training schedule of these two levels. Experiments use self-distillation with a frozen teacher across three models and four interactive environments. Results show that student-induced histories with teacher-generated action prefixes generally give the strongest standalone signal, while following the student at both levels performs worst.
In briefPermission to Retry: Study of how AI agents recover safely after uncertain actions
SummaryThis paper examines how tool-using AI agents recover from interrupted actions, separating recovery decisions from language-model behavior. The authors test checkpoint state, dispatch permission, evidence of external effects and reconciliation permission across 336 histories on LangGraph and Temporal. Retained-ledger reconciliation passed all 42 histories per host, while cached retry, revalidation before retry and stable-key retry passed 30, 36 and 31 respectively.
In briefDecode Late, Multiply Once: A New Way to Compute Dot Products with Quantized Weights
SummaryThis paper targets neural networks whose weights use only 1 to 4 bits, proposing a dot-product method where each weight code selects one of 2^k accumulators and activations are added into them. The accumulated sums are decoded and multiplied just once at the end, so the decoded weights are never formed. The authors report that the method is exact for integer arithmetic and matches the conventional pipeline's floating-point error, and an FPGA study of a scatter-accumulate processing element shows area savings only when the multiplier is shared across PEs.
In briefLazyFlash: Dense Forward, Sparse Backward Attention to Speed Up Long-Context Fine-Tuning
SummaryThis paper introduces LazyFlash, an attention method that keeps the forward pass dense while making the backward pass sparse, to speed up fine-tuning of large language models on long contexts. It ranks query blocks by attention-output gradient norms and skips those that contribute little to parameter updates, implemented inside the FlashAttention backward kernel. Experiments on long-context text and speech LLMs report up to 5.7x attention-backward speedup and 1.7x end-to-end training speedup while maintaining downstream performance.
In briefShows spectral and density clustering hierarchies are topologically equivalent
SummaryThe paper compares two ways to build hierarchical clusters from a density: density cluster trees from connected superlevel sets, and spectral hierarchies from single linkage on Laplacian diffusion distances. It gives conditions under which, at a suitable diffusion time, the two are topologically equivalent up to monotone rescaling of heights. It also shows a k-NN graph implementation inherits this equivalence via uniform approximation, reducing finite-sample spectral constructions to density level-set topology.
In briefPixGR: Pixel-space generative decoding for fast 3D scene reconstruction from sparse views
SummaryThis paper presents PixGR, a framework that adds a lightweight pixel-space generative model to feed-forward 3D Gaussian reconstruction to account for scene uncertainty. Context views are encoded once into a scene latent, which conditions generation of a shared 3D Gaussian representation from noisy target views. Training uses pixel-space flow matching with differentiable rendering, avoiding VAE latent compression. In experiments, PixGR outperforms regression-based methods in sparse-view settings with widely spaced or extrapolated views while keeping inference under one second.]
In briefCalibrating Bosch etching simulators to real measurements using few XSEM samples
SummaryThis paper proposes a few-shot real-to-sim calibration framework for Bosch etching, a process used in MEMS and TSV fabrication. Using a small set of real XSEM measurements, it infers a recipe-specific ViennaPS parameter vector that reduces the gap between simulated and measured etch depth and sidewall angle. The search combines local exploration, LLM-guided candidate generation, and XGBoost surrogate screening, with ViennaPS simulation used for final verification. Experiments on two recipes predict held-out etching results with normalized global errors of 0.340 and 0.249.
In briefAutomated nucleus segmentation in histology images using an Attention U-Net pipeline
SummaryThis paper addresses nucleus segmentation in digital pathology, where dense nuclei and staining variation make analysis difficult. It uses a lightweight Attention U-Net on the MoNuSeg 2018 dataset, with stain deconvolution, mask smoothing, and overlapping patch stitching for high-resolution inference. The model reached an IoU of 0.8745 on unseen full-resolution test images, which the authors present as evidence for scalable automated segmentation.
In briefDistributed Swarm Shape Building with Neural Cellular Automata: Robots Grow Target Shapes Through Local Interactions
SummaryThis paper presents a modular framework that combines neural cellular automata (NCA) with site assignment and collision-aware control so robots can assemble desired shapes through local interactions. It introduces a decentralized mechanism in which robots already in the formation share local NCA states to find new attachment sites, allowing growth without a predefined site list. Offline sparse-retention fine-tuning and a false-positive loss are proposed to improve morphology accuracy. Simulations show diverse lattice shapes with up to several hundred robots, and experiments with Crazyflie quadrotors validate the execution pipeline.
In briefStateBridge: Phase- and track-based reasoning for captioning and answering questions about traffic videos
SummaryThis paper presents StateBridge, a framework trained only on synthetic traffic videos that generates captions and answers visual questions about real-world traffic scenes. It uses phase-aware clip construction, pedestrian and vehicle track summaries, and structured prompts to ground vision-language models in stable traffic evidence. The authors report an official S2 score of 55.4679 on AI City Challenge 2026 Track 2.
Blessing Agyei Kyem, Richard Dzinyela, Joshua Kofi Asamoah et al.
In briefLearning structured sparsity for large models to link compressibility with efficient inference
SummaryThe paper introduces S-SALAAD, a framework that jointly learns model weights with low-rank and structured sparse components during training. It adds budget-aware compression and dedicated CPU and GPU kernels so a single checkpoint can run efficiently under different memory budgets. Experiments on Llama2 and Qwen3 models, covering pretraining and distillation, report improvements in perplexity, memory, and inference latency over the evaluated baselines across several hardware platforms.
In briefKnowing When to Stop: The Evidence-Action Gap in Coding Agents
SummaryThe paper finds that coding agents keep trying to finish tasks even when the environment makes completion impossible, rarely acting on early evidence that only outside help can restore progress. The authors introduce ImpasseBench, which injects twelve environment blockers into solvable repair tasks and measures actions against the step where evidence becomes sufficient. Across eight frontier models and 19,371 trajectories, nearly all runs needing outside help collected sufficient evidence, yet a third exhausted their budget, and an escalation tool reduced budget exhaustion from 34.0% to 6.7%.
Tianyu Chen, Xinran Zhao, Vijay Viswanathan et al.
In briefTaxoSAE: Rethinking Sparse Autoencoders with Taxonomic Logic
SummarySparse autoencoders (SAEs) break down large language model representations into interpretable features that can be read out and controlled. This paper notes that existing SAEs treat each feature independently, even though real-world knowledge is often organized hierarchically through is-a relations. The authors propose TaxoSAE, which stacks one SAE per taxonomy level and uses a taxonomic alignment loss so a concept activates whenever one of its hyponyms does. Experiments on two factual QA taxonomies report gains in readout accuracy, control precision, and reading out unseen concepts through their hypernyms.
Xinyu Guo, Minglai Yang, Clayton T. Morrison et al.
In briefVisFlow: Interpretable Flow Matching for Reconstructing and Characterizing Polygon Visibility
SummaryThe paper presents VisFlow, an interpretable flow formulation that combines continuous signed-distance-function (SDF) representations with explicit polygon evolution. It addresses recovering polygon geometry from visibility graphs, a task where small shape changes can abruptly alter the discrete visibility structure. VisFlow is evaluated on Visibility Reconstruction and Visibility Characterization. According to the paper, it maintains 20% more visibility relations on reconstruction and improves coverage by 30% on characterization compared with state-of-the-art methods.
In briefCombinatorial conditions for learning conditional distributions from data
SummaryThis paper studies PAC-learnability of function classes that map inputs to probability distributions over a finite space, where labels are drawn from an unknown target's conditional distribution. Learning is measured by excess risk under a bounded proper loss. The authors introduce the fat_γNatarajan and fat_γGraph dimensions and show that finiteness of the fat_γNatarajan dimension at every positive scale is necessary and sufficient for learnability. They further show a distribution-valued class is learnable if and only if each coordinate class is learnable in the probabilistic concept model.
In briefScaling Transformer energy-based models for high-quality image synthesis
SummaryThis paper introduces the Transformer Energy-based Model (TEM), a scalable transformer energy function operating in latent space, which sums energies across multiple layers to combine local and global constraints. Positive Representation Alignment supervises intermediate representations with a vision foundation model, and advanced sampling strategies improve quality at inference. The reported results include an FID-50K of 3.82 for TEM-XL on class-conditional ImageNet 256×256 without classifier-free guidance, and 8.24 at 512×512.
WhyScales energy-based image generation to ImageNet 512 with a transformer energy and representation alignment, reaching FID 3.82 at 256, with solid experimental support.
In briefCurriculum learning works best when rewards are sparse and optima compete.
SummaryThis paper builds a minimal policy-learning model to study why curriculum learning helps more in reinforcement learning than in supervised learning. It analyzes an open-loop, episodic target interception task whose reward sparsity can be adjusted, using both mathematical and numerical methods. With dense rewards, curricula offer only marginal benefit, while with sparse rewards they can be required for learning to occur at all. The effect is strongest when the agent must balance minimizing effort against accumulating high reward.
John J. Vastola, Ann Huang, Satpreet H. Singh et al.
In briefTACS-RF: A Tool-Driven AI Agent Framework for Designing RF and Millimeter-Wave Circuits
SummaryThe paper presents TACS-RF, a framework in which a large-language-model coding agent directly operates industry-standard electronic design automation tools to design RF and mmWave modules. The authors validated it by semi-autonomously designing a 240–251 GHz sub-THz transmitter in a 90 nm SiGe BiCMOS process, including oscillators, power drivers, a frequency doubler, an antenna, and matching networks. The transmitter delivers +5.1 to +6.3 dBm output power, phase noise of -80 to -88 dBc/Hz at 1 MHz offset, and consumes 361 mW of DC power.
In briefFSVL: A fast-slow vision-language framework for identifying confusable Yangtze fish species and generating expert knowledge
SummaryThe paper proposes FSVL, a two-pathway system that pairs image recognition with a retrieval-augmented large language model to identify 51 fish species in the Middle Yangtze River and generate species knowledge. A fast ResNet50 classifier performs initial identification, and a slow path uses YOLOv8n to examine key body regions for confusable species. On a 607-image test set, the fast classifier reached 99.67% accuracy and the cascade reached 100.00%. On 255 knowledge generation samples, scientific name accuracy rose from 91.37% to 100.00%.
In briefMining Frames: Unsupervised Association Rule Mining for Detecting News Framing
SummaryThis paper proposes association rule mining as an unsupervised way to detect news framing without manual annotation or predefined categories. The main validation covers 25,512 English-language articles from 16 US, UK, French, and German outlets on the first year of the Gaza war. High-lift word co-occurrence rules reveal interpretable framing patterns, which journalist-researchers confirmed in context. The study also finds systematic differences between headlines and body text, and two published investigative stories drew on the findings.
Hager Hesham Mustafa (Anmat Computational Analysis Lab)
In briefForecasting Associations, Not Interventions: A Causal Audit of Time Series Foundation Models
SummaryThis paper examines whether covariate-aware time series foundation models answer what-if queries with the conditional slope or the interventional effect of a control variable. The authors derive a parameter-free reference line for the confounding gap and test several models on a preregistered grid with exact counterfactual ground truth. They find the models respond more confounded than regression, and that brief interventional post-training on synthetic series reduces this gap by up to 83%.
WhyDerives a closed-form confounding gap for observational predictors answering interventional queries, validated by preregistered audits, real actuator experiments, and a post-training fix.
In briefLunar robot swarms learn to share warmth: emergent cooperative thermoregulation
SummaryThis study examines whether small robots surveying lunar caves can survive the 354-hour night by using each other as thermal shielding. One policy shared by twenty robots was trained with reinforcement learning, receiving a single reward at night's end based only on the mean temperature of the five coldest robots. The swarm aggregated and rotated between sheltered and exposed positions, needing 37% less stored energy than a motionless huddle, and the behavior transferred to forty robots without retraining.
In briefScaling laws for training on synthetic rewrites under compute and data limits
SummaryThe paper studies synthetic rewrites, where a language model rewrites natural documents into a target style for pretraining. The authors propose a scaling law that predicts loss from model size, training tokens, rewrite fraction, and natural-data epochs, including the compute spent generating rewrites. In a compute-limited setting, the estimated optimal rewrite fraction rises with compute, and a 22B-parameter model trained this way beats its rewriter on several web domains. In a data-limited setting, a model trained on rewrites of its own data outperforms natural-data-only training up to 3B parameters.
WhyProposes a rewriting scaling law charging rewrite compute, validated in compute- and data-limited settings with accurate 33x FLOP extrapolation.
In briefMachine-readable operator manuals over ISO 11783 for offline, multilingual LLM assistance
SummaryThis paper studies how a tractor-side assistant's local knowledge base can be synchronized with the operator manual of a connected implement over an ISO 11783 CAN bus. The authors benchmark Markdown, JSON, and XML transfer and evaluate open-weight language models on cross-lingual questions. A complete manual transferred in 31 to 89 seconds, and locally run models answered most English questions correctly but performed worse on the other five language editions. Semantic retrieval supported cross-lingual queries, while keyword retrieval did not.
In briefAligning language models with human feedback: math and algorithms
SummaryThis tutorial-style paper explains the mathematical foundations and algorithms used to align large language models with human intentions, preferences, and values. It covers supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and direct preference optimization (DPO). It also connects learning from preferences to inverse reinforcement learning and discrete choice models, and discusses the advantages, limitations, and practical implementation of these methods.
In briefCautious Conflict-Free Inverse Gradient Methods: Fixing Gradient Conflicts in Neural PDE Solvers
SummaryPhysics-Informed Neural Networks train on several loss terms that often push gradients in conflicting directions. The paper shows that combining the ConFIG conflict-free method with adaptive optimizers like Adam can break that conflict-freeness at the weight update. The authors propose Cautious ConFIG, a three-line extension that restores conflict-freeness before the update. Experiments on three PINN benchmarks report consistent gains over standard ConFIG.
In briefSWE-Clar: Can AI Coding Agents Clarify Vague Client Requests and Build Software from Scratch?
SummaryThis paper introduces SWE-Clar, a benchmark of 100 software tasks drawn from real Upwork jobs, each paying at least US$4,000. Coding agents first hold multi-round conversations with a simulated client, then are evaluated separately on code implementation with acceptance tests and on 512 requirement-understanding questions. Across nine frontier models, the best one passes only 67% of tasks.
In briefRethinking Multi-View Fusion for Urban Representation Learning: Introducing the PRISM Adaptive Framework
SummaryThis paper studies how to fuse heterogeneous urban signals, such as mobility flows, points of interest, and land-use patterns, into unified region embeddings for tasks like check-in prediction, crime forecasting, and service demand estimation. The authors argue that existing globally shared fusion and joint optimization overlook differences in modality importance across regions and cause gradient imbalance among objectives. They propose PRISM, which uses sparse MoE routing to learn region-specific view allocations and a PID-inspired mechanism to mitigate gradient conflicts. Experiments on three large-scale urban datasets report competitive performance against strong multi-view baselines, with more stable optimization and more balanced region representations.
In briefIFC-SSL-Bench: Benchmarking self-supervised learning to reduce labeling needs in multispectral flow cytometry imaging
SummaryThis paper presents IFC-SSL-Bench, a modular framework for evaluating self-supervised learning (SSL) on imaging flow cytometry data. The authors benchmark six SSL approaches across four backbones on two datasets, measuring classification performance across annotation levels from 1% to 100%. They report that SSL matches fully supervised baselines using as little as 5% of annotations for morphological classification, with MoCo-v2 and ConvNeXt backbones performing best.
Angelo Jovin Yamachui Sitcheu, Roxana-Teodora Fetescu, Thomas Lautenschlager et al.
In briefDon't Forget What You Just Learned: A method that turns agent mistakes into lasting, reusable lessons
SummaryThis paper proposes Self-Direction, a mechanism in which an agent reflects on a failed task trajectory, extracts a lesson, and distills it into a lightweight low-rank adapter while keeping base weights frozen. The adapters are stored in a persistent library and automatically routed to related problems later. Experiments with Qwen-family open-weight models on GAIA, Claw-Eval, and Terminal-Bench 2.1 report success-rate gains of 8.9 to 17.0 percentage points and fewer repeated mistakes.
Yuval Kansal, Karen Ullrich, Polina Kirichenko et al.
In briefEgoForesight: Scaling Robot Policy Learning with Noisy Egocentric Actions
SummaryThis paper presents EgoForesight, a vision-language-action framework that learns from both robot demonstrations and egocentric videos. It reduces the impact of unreliable action labels and uses demonstrated outcomes as extra supervision through future visual prediction and teacher-student distillation. According to the paper, over the tested range it shows steeper scaling than the baseline, cutting real-world open-loop action MSE by 20.5% and raising RoboCasa few-shot success by 8.1 points at 30,000 hours of pretraining data.
In briefScaling cross-situational word learning to children's real-life videos
SummaryThis paper studies whether children learn early concrete nouns through cross-situational learning, using contrastive vision-language models trained on the BabyView corpus of 2,441 hours of egocentric video from 48 children. Models using internet-pretrained visual representations performed far better than those learning vision solely from the developmental video data. Most of the learning signal came from a small share of utterances describing a visible referent. The learned word representations showed human-like semantic structure, but performance trajectories over training differed from those of children.
Michael C. Frank, Alvin W. M. Tan, Khai Loong Aw et al.
In briefSpecMem: Speeding Up Tool-Calling AI Agents with Live Memory
SummaryTool-calling AI agents are slowed by the need to decode every tool call step by step. This paper introduces SpecMem, which keeps a memory of verified tool calls that is updated throughout deployment, with per-user bounded stores and LRU eviction, and defers irreversible calls until verification. In tests across multiple sessions and users with three open-weight models, it achieves a 27% higher mean accepted token count than a frozen store and up to 1.74x wall-clock speedup.
In briefSAGE: A Benchmark for Multimodal Word-Image Grounding with Gloss-Based Reasoning
SummaryThis paper presents SAGE, a unified 10-candidate image-selection benchmark that merges VWSD, VerSe, and WikiDiverse into one term-centered protocol covering nouns, verbs, and entities. It includes confusable hard negatives and a context-swapped VerSe diagnostic. The authors also propose GAR, a training-free method combining definition grounding with constraint-guided semantic verification. Combining both signals raises macro-average Hits@1 from 65.90 to 71.22.
In briefSAGE: A hybrid vision-language and geometry framework for predicting vehicle paths through work zones under temporary traffic control
SummaryThe paper proposes SAGE, a framework that helps automated vehicles predict which path to follow through road work zones with temporary traffic control. It separates generating geometric candidate paths, choosing among them, and checking compliance with traffic control rules, using a vision-language model's trajectory to rank candidates. Tested on 2,306 real-world ROADWork validation episodes, SAGE reduced displacement errors relative to a geometric ranker and cut hard-constraint violations by 37%.
In briefWorldMirror: An agent framework that builds interactive simulated worlds from real robot demonstrations
SummaryWorldMirror is an agent-based framework that turns real-world robot demonstrations into interactive digital twins for simulation. It treats reconstruction as iterative refinement of an executable world program, with a coding agent calling image and 3D generation tools and correcting errors based on verification feedback. Experiments on the DROID dataset show it can reconstruct visually faithful, geometrically consistent worlds while preserving the physical validity of recorded robot trajectories.
In briefLatent Coherence: Whether Frontier Language Models' Probabilities Follow the Total Probability Law
SummaryThe paper tests whether frontier language models, without special training, give probability judgments consistent with the law of total probability when a latent variable must be inferred from context records. Models are asked separately for P(X), P(z), and P(X|z), and a causal version of the test is also applied, with exact ground-truth probabilities for every query. Consistency differs clearly across seven models, and larger inconsistency residuals correlate with larger errors against the truth, so the residual can flag unreliable answers without ground truth.
In briefSEVI: A safety method for real-robot learning that overrides risky actions
SummarySEVI is a framework that makes real robots safer while they learn new tasks online. Its Evidential Safety Critic estimates future constraint-violation risk and epistemic uncertainty in a single forward pass, and unsafe or poorly supported actions are replaced by an offline-trained recovery policy. Tested on two simulated manipulation tasks and a real USB insertion task, SEVI reports the best safety prediction, task success, and fewest unsafe events.
In briefWhat the Query Remembers: How Query Abstraction Shapes Lexical and Dense Search of Classical Arabic Poetry
SummaryThis paper benchmarks nine dense encoders and BM25 on Classical Arabic poetry retrieval across two corpora, three indexing modes and two evaluation protocols. Which retriever performs best depends on how much of a poem's original wording the query still carries. BM25 beats every dense encoder on tip-of-the-tongue queries but drops to eighth of ten on abstract queries. The authors argue that a benchmark reporting only one protocol or one query type can give a misleading picture.
Omar Farouk Zouak, Selma Mani, Larbi Said Chikh et al.
In briefA free confidence estimator for LLMs: reusing the RL critic
SummaryThis paper studies confidence estimation for large language models. The authors prove that, under a correctness-based outcome, confidence estimation and value-function (critic) training in reinforcement learning are equivalent. As a result, the existing critic of a value-based RL-trained LLM can be reused as a confidence estimator without extra training or data, leaving the policy's task performance unchanged. Experiments across math, code, tool calling, and general reasoning report that the reused critic is the best estimator on every in-distribution split and the best-calibrated on 14 of 16 settings.
WhyProves confidence estimation equals critic training under correctness outcomes, reusing existing critics; supported by experiments across many domains.
In briefSpeculative Diffusion Balances Two Tradeoffs: Tuning Diffusion Effort and Acceptance Leniency for Faster Language Model Inference
SummaryThis paper studies speculative decoding, which speeds up language models by using a faster draft model, with diffusion models serving as the drafter. It varies two parameters together: diffusion effort and the leniency of the acceptance criteria, which can trade output quality for speed. The authors map the resulting speed-quality Pareto frontier. Their experiments show that jointly varying these parameters extends the frontier beyond what single-step diffusion drafting achieves.
Aaron Zorel, Guy Van den Broeck, Aditya Grover et al.
In briefNested sparsity: a new conditional computation method for scalable pretraining and continual learning
SummaryThe paper proposes a conditional computation mechanism that generalizes the attention operator, with sparsity that nests within the sparsity of mixture-of-experts (MoE) transformers; the resulting model is called Nestor. In language model pretraining, at equal compute and sparsity, Nestor outperforms a conventional MoE transformer. For continual pretraining, it adapts more effectively to longer contexts and new domains while forgetting less on old domains. These benefits persist as model size, dataset size, and context length grow. Code and checkpoints are available on GitHub.
WhyExtends conditional computation to attention and nests it with MoE sparsity; scaled experiments on pretraining and continual learning show consistent gains over MoE baselines.
In briefPrefO2O: Zero-shot preference-based reinforcement learning via offline-to-online latent preference transfer
SummaryThis paper introduces PrefO2O, a framework for teaching robots behaviors aligned with human preferences without new human feedback. It learns a task-invariant latent preference space from a small amount of offline preference data on source tasks, then uses confidence-weighted optimal transport to create pseudo-preference supervision for adapting rewards on new tasks. Experiments on robotic manipulation and locomotion benchmarks report that it outperforms preference-transfer baselines and matches or exceeds methods trained with real human feedback.
In briefPaperFormalizer: Author-guided conversion of research papers into Lean-verified form
SummaryThis paper presents PaperFormalizer, a Lean-based agentic pipeline that helps authors find and fix small gaps in the proofs and assumptions of their own papers and convert each paper into a Lean-verified version. It starts from the LaTeX source, audits each statement for faithfulness, ambiguity and correctness, and asks the author whenever a statement has more than one reading. Every proposed repair is left for the author to accept, decline or redirect. The output is a revised paper, a change report, and Lean code linked to the statements it formalizes.
In briefCuraLoc: Planning code inspection to locate bugs across a repository
SummaryThis paper presents CuraLoc, a method for finding which repository code to change given a bug report. It first selects candidate files and describes their symbols compactly, then uses one model planning step to choose starting symbols, relationships to inspect, and likely edit hypotheses. Deterministic retrieval gathers the requested source for a final ranking. The authors report strong accuracy on SWE-Bench Lite and Verified, with 90.3% fewer recorded tokens than LocAgent on Verified using GPT-5.2.
In briefDetecting Patronizing and Condescending Language in English and Mandarin
SummaryThis paper examines patronizing and condescending language (PCL), which belittles vulnerable groups while assuming pity, reinforcing negative stereotypes, and is often hard to detect even for human judges. The authors propose an automated detection system with two approaches: static word embeddings with a convolutional neural network, and pre-trained BERT embeddings fine-tuned for the task. The BERT-based approach was considerably more successful, achieving an F1 score of 0.528 on the development set.
Xinyi Vicky Xiang, Catherine Qingyun Wang, Paula Cort\'es Lemos
In briefStackGen3D: Generating Stable 3D Block Stacks from Multiview Silhouettes
SummaryThis paper presents StackGen3D, a framework that generates physically stable block assemblies from a block inventory and three orthographic silhouettes. It uses a physics-informed graph neural network to predict support relations and curate a stable-structure dataset, then trains a conditional flow-matching model to produce configurations matching the silhouettes. Best-of-10 selection raised physical stability from 45.3% to 72.4%, with a silhouette IoU of 0.970.
In briefPnP*: A certifiable solver that finds globally optimal camera poses from 3D-2D matches
SummaryThe paper addresses camera pose estimation from 3D-2D point correspondences, known as Perspective-n-Point, by targeting the true maximum likelihood objective. It proposes PnP*, presented as the first scalable solver that returns a certifiably globally optimal pose. The authors prove the underlying convex relaxation is exact under moderate noise and reduce it using chordal sparsity and quotient ring reduction. Numerical experiments verify stability and scalability, with the solver running at over 10 Hz.
WhyFirst scalable certifiable solver for PnP maximum likelihood with proven exact convex relaxation under moderate noise, numerically verified; useful for visual localization.
In briefPaperScorer: Predicting whether a paper gets accepted from its content alone
SummaryThis paper introduces PaperScorer, models that predict a paper's accept or reject decision from its extracted text or page screenshots, with author identity and formatting cues removed. Vision models performed best, reaching 68% accuracy on ICLR papers and 77% on arXiv ML papers, beating text-only and frontier reasoning models. Predictions were weakest for mid-rated papers, and the model's most confident predictions were also its most accurate.
In briefPhronix: A Capability-Centric Learning Paradigm for AI Agents
SummaryThis paper introduces Phronix, a learning approach that lets AI agents acquire, validate, and refine reusable capabilities from their own executions through a predictor-validator-corrector loop. The authors report that learned capabilities solve recurring tasks on the first attempt in 93-95% of cases across models. On a 1000 Genome Pegasus workflow, Phronix passed more correctness checks than the baseline harness.
Hariharan Devarajan, Neeraj Rajesh, Aniruddha Marathe et al.
In briefNot All Retained Clients Are Equal: Client Selection for Federated Unlearning
SummaryThis paper addresses federated unlearning, which removes a target client's influence from a trained model without full retraining. It proposes PDF, which records each client's path-dependent update history during training and, upon an unlearning request, selects the top-k retained clients that are globally informative yet weakly coupled to the target for brief retraining. Experiments on FMNIST, CIFAR-10, and CIFAR-100 report effective forgetting and strong retained utility with far fewer clients, with communication reduced by about 69% and computation reduced relative to Retrain.
In briefEarly learning in large residual networks: a study showing that early training in very high input dimensions only learns linear functions, suggesting some mean-field models may be too rigid to explain feature learning.
SummaryThe paper studies early training dynamics of nonlinear residual networks built from blocks with zero hidden layers, using a sequential limit where width and depth, then training set size, then input dimension, diverge. The authors compute network outputs and hidden representations after a few gradient descent steps diagrammatically, in terms of statistics of inner products of training inputs. A main negative finding is that, for large input dimension, early training can only learn a linear function, as higher monomials are suppressed and vanish. The paper also examines how to keep the learning rate's hyperparameter transfer consistent across dataset size and input dimension, and reports numerical tests in strong agreement with the theory.
In briefBenchmarks Are Surprisingly Compaction-Friendly
SummaryLong-horizon AI agents must free up their context windows through compaction, which is usually lossy. The authors implemented five compaction strategies in a fixed agent scaffold and evaluated three frontier models on three long-horizon benchmarks. They found that compaction threshold had little effect for large-context models, but compaction can silently drop task constraints, with losses rising as compactions accumulate.
Jeanne Salle, Hardik Bhatnagar, Matthias Bethge et al.
In briefCutting Omni-model inference cost by choosing modalities per query
SummaryOmni-modal large language models can process text, audio, and video together, but this is costly at inference time. The paper observes that extra modalities help different queries by different amounts, so it keeps the model fixed and adaptively selects which modalities to provide for each query, framed as an accuracy-cost optimization. Tested on WorldSense with Qwen2.5-Omni and Qwen3-Omni, the approach improves the accuracy-cost trade-off. For Qwen2.5-Omni it matches or exceeds full-modality accuracy using about 41% of input tokens, 39% of latency, and 61% of monetary cost.
In briefSyncVid: Synchronized multi-view video generation from one video
SummarySyncVid generates multiple synchronized videos of a dynamic scene from a single reference video, keeping content consistent across viewpoints and over time. Instead of sampling noise independently in each 2D image plane, it lifts the source video into a canonical 4D representation, samples structured noise in 3D space, and rasterizes it into target cameras. A cross-view-consistent whitening step decorrelates neighboring pixels, and the structured noise serves as the prior for a LoRA-adapted pretrained video diffusion model without architectural changes. The authors report improved perceptual quality and cross-view consistency.
Farhad G. Zanjani, Animesh Karnewar, Fatih Porikli et al.
In briefVibeGame: A framework that builds 2D games from prompts using an AI-native engine and a self-evolving team of adversarial agents
SummaryThe paper presents VibeGame, which creates complete 2D game projects from natural-language prompts and optional reference images. Its AI-native engine stores project state as structured text and offers frame-synchronous runtime control so agents can inspect, act, and verify game behavior. Eight specialized agents separate implementation from review and fix problems through multi-round loops, and accepted projects are distilled into reusable skeletons and modules. On a GameCraft-Bench subset, the prompt-to-game setting consistently outperformed other agent harnesses using the same underlying model.
In briefNavigation model that learns its own visual representation and predicts future scenes
SummaryThe paper presents RepresentNav, a navigation foundation model that builds its own representation space from multiple vision foundation models, so the policy learns to see what navigation needs. Before each action, it predicts a latent representation of the scene a moment ahead, trained end-to-end as a joint-embedding predictive architecture. The authors report new state-of-the-art results on several navigation benchmarks and gains on a physical robot.
WhyProposes a self-constructed representation space with JEPA-style latent dreaming, showing large gains across several navigation benchmarks with fairly thorough evidence.
In briefAudioLift: Audio-driven full-head 3D avatars from one photo via generated multi-view video
SummaryAudioLift is a two-stage framework that creates a 3D talking avatar from a single portrait image and audio. It uses generated multi-view videos as an intermediate representation, avoiding parametric face models and per-instance optimization. A robust Gaussian reconstruction stage recovers coherent full-head geometry. The authors report it captures more extreme expressions than prior methods, with better geometric completeness and multi-view consistency across 360 degrees.
In briefA Unified Model for Image-to-3D Part Generation at Variable Levels of Control
SummaryThis paper presents FlexPart, a single model that generates 3D object parts from an image while accepting geometric conditions of different granularity, from sparse points to bounding boxes and masks. It fuses these heterogeneous inputs through gated adaptive modulation and adds an asymmetric geometric guidance mechanism to improve geometric consistency. Under mask conditioning on PartVerseXL, the authors report Part CD reductions of 16.6% relative to TripoSG + SAM3D and 10.5% relative to OmniPart.
In briefSensorimotor Alignment: A Low-Cost Offline Proxy for Predicting Robot Task Success
SummaryThe paper proposes Sensorimotor Alignment (S²), a training-free offline metric that measures how well a model's sensory representations match the geometry of physical motion. S² correlates positively with downstream robot task success across model families, so it can serve as a cheap proxy for ranking models before closed-loop evaluation. It can also help select vision-language model backbones before fine-tuning into vision-language-action models, and offers representation-level explanations for some observed behaviors.
In briefReading, writing, sharing, and interacting with the world through video: a world-modeling survey
SummaryThis survey reviews video foundation models and video world models through a unified world-modeling lens. It treats observed video as a partial, lossy, temporally ordered projection of a latent world and organizes work around four operations: reading, writing, sharing, and interacting. The authors note a shift in emphasis from clip-level recognition and perceptual fidelity toward persistent state, controllable evolution, physical and causal validity, and interaction-based evaluation, and they outline the open challenges that follow.
In briefDirect Pass@k Optimization for Pruning Reasoning Paths: Balancing Quality and Diversity at Test Time
SummaryThis paper addresses test-time scaling for reasoning models, where many candidate solutions are generated and then pruned after a short prefix to save computation. It proposes a pruning method that directly optimizes pass@k over the selected pool, maximizing the chance that at least one partial solution leads to a correct answer. This criterion naturally balances quality and diversity. Across eight open-weight reasoning models, the method outperforms existing quality-based and diversity-based pruning strategies, with the largest gains when a moderate number of partial solutions is kept.
Mahdi Salmani, Adel Javanmard, Devansh Gupta et al.
In briefPhase transition from random features to fully trainable networks
SummaryThis paper studies how partially trainable neural networks, which combine fixed random features with adaptive components, approximate target functions. The authors derive a squared L2 approximation rate for two-layer networks under polynomial spectral decay assumptions and report that training errors show a similar phase transition: rapid, smooth early gains followed by diminishing returns. Numerical experiments with networks of up to 15 layers are included.
In briefReliable AI decisions through selective expert auditing
SummaryThe paper studies online monitoring of an AI system that outputs decisions with confidence scores, deciding whether to accept each decision automatically or send it to a human expert. It formulates this as an online constrained optimization problem that minimizes expert auditing frequency while keeping the fraction of unaudited AI errors below a user-defined threshold. The authors derive algorithms with provable error-control guarantees under no assumptions, including one that asymptotically achieves the best auditing rate for both adversarial and i.i.d. streams. Experiments on screening harmful requests to large language models, text annotation, and data labeling show valid error control and substantially reduced human auditing.
WhyFormulates selective auditing as online constrained optimization with assumption-free error-control guarantees and an asymptotically optimal method, validated on LLM screening and annotation tasks.
Isaac Gibbs, Elise L. Han, Ryan J. Tibshirani et al.
In briefSpectralCanary: Low-Frequency Signal Traces for Auditing Text-to-Image Dataset Use
SummaryThis paper proposes SpectralCanary, which blends a fixed canary image into every dataset image using a smooth low-pass mask before release, leaving captions unchanged. Auditors then check whether this signal appears in a suspect text-to-image model's generated outputs to detect unauthorized fine-tuning on the collection. Reported experiments across several datasets and two model families show high detection rates with minimal change to released images, and matched-energy ablations favor low-frequency placement over mid or high frequencies.
Sy-Tuyen Ho, Anirudh Satheesh, Soham Chawathe et al.
In briefCost-aware abstention for automatically routing IT support tickets using LLMs
SummaryThis paper studies how to decide when an automated system should route IT support tickets or defer them to human triage, given the costs of wrong routing. The authors select a routing policy on calibration data under explicit costs, then evaluate it once on held-out test data. On a public corpus of 47,837 tickets across eight classes, they compare four trained non-LLM classifiers with four large language models, reporting that a TF-IDF with logistic regression policy reduced test cost under the illustrative cost model. The findings are scoped to this dataset, prompts, model endpoints, and illustrative costs.
In briefMETTA: Memory-enhanced test-time adaptation for meta-reinforcement learning
SummaryThis paper proposes METTA, a meta-reinforcement learning method that helps agents adapt quickly to unseen tasks from limited interaction. It splits adaptation into two timescales: a causal neural memory handles fine-grained adaptation within an episode, while compact episode-level summaries support long-range task inference across episodes. The authors report strong success on unseen Meta-World ML10/45 test environments and better structured exploration than transformer baselines on sparse-reward PointMaze tasks.
Ali Najar, Hosein Hasani, Amir Kooshan Fattah et al.
In briefBrickWright: AI agents building real, physically valid LEGO models
SummaryThis paper presents BrickWright, a framework that lets pre-trained coding agents generate LEGO brick builds from text or image prompts. The agents write code using BrickLayer, a Python toolkit that provides a brick assembly API and a validator checking that parts are real, do not collide or float, and form a connected structure. According to the paper, baseline agents reached 9% to 69% physical validity, while BrickWright-guided agents reached 100% on a sample of 512 prompts.
In briefEpiSEM-Bench: Testing whether LLM agents stay safe when key evidence is missing
SummaryThis paper introduces EpiSEM-Bench, an executable benchmark that tests whether tool-using LLM agents take unsafe actions, respect safety boundaries, and recover missing evidence. It contains 300 matched unsafe and safe task pairs across six domains, and suppresses critical facts to measure failures. Results across ten models show large differences in robustness to omission, in actively acquiring missing facts, and in acting appropriately.
In briefGate3DGen: a preprocessing model that selectively cleans real-world images for 3D generation
SummaryGate3DGen is a preprocessing model that sits in front of existing 3D generators without retraining them. A degradation-aware router passes clean images through unchanged and sends degraded ones to a one-step purifier that restores them while preserving object identity, contours, and texture. On a synthetic benchmark it takes 2.3 seconds on average and reduces TRELLIS.2 mesh distance by 61.5% relative to degraded inputs.
In briefLow-bit n-gram embedding tables cut memory while keeping language model gains
SummaryThis paper studies storing n-gram embedding tables used in language models at low precision to reduce their memory footprint. The authors compare different quantization methods and bit allocations. On models trained from scratch, four-bit tables save about 75% of table memory while keeping over 99% of the validation-loss improvement from the tables. On large existing models such as DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next, low-bit tables save 47-55% of table memory with perplexity increases under 1.2%.
In briefAgentic AI for Public Health Data Collection
SummaryThe paper proposes a multi-agent framework that coordinates disease surveillance data collection across heterogeneous sources through source discovery, assessment, evidence structuring, event linking, and validation. Three retrospective case studies are presented: it recovered all 13 individual cases in a hantavirus outbreak, reconstructed Canadian measles counts within 0.33% of the national reference total, and raised the share of days with recent Sierra Leone mpox reports from 28.2% to 65.6%. The authors note limitations in record completeness and accuracy and state that prospective gains in timeliness remain untested.
In briefOmniRay: Camera-Agnostic Monocular 3D Reconstruction via Composite Ray Fields
SummaryThis paper presents OmniRay, a feed-forward model that recovers 3D geometry from a single image without needing camera information at inference time. Its Ray Module combines a monotone radial base, a spherical-harmonic residual, and a free-form residual to represent diverse camera geometries such as fisheye and panoramic lenses. The authors report that experiments across different camera configurations establish a new state of the art in camera-agnostic monocular 3D reconstruction.
Namitha Guruprasad, Abhay Yadav, Cheng Peng et al.
In briefGigaBench: A Benchmark for Learnable Representations in Gigapixel Image Compression and Super-Resolution
SummaryThis paper introduces GigaBench, the first benchmark for testing learnable representations (LRs), which fit parametric functions to signals, on gigapixel images. It includes images from diverse domains and covers compression and super-resolution tasks. The authors adapt implicit, hybrid, and explicit LR methods to run at gigapixel scale on consumer hardware and evaluate reconstruction quality, model size, and compression and inference times. The study reports consistent trends across methods and offers guidance for future research.
Matteo Fusconi, Alex Costanzino, Pierluigi Zama Ramirez et al.
In briefA fix for two numerical bugs in softmax and cross-entropy
SummaryThis paper identifies two numerical bugs in standard softmax, log-softmax, and cross-entropy implementations. One, Softmax Collapse, rounds the target logit's gradient to zero when that logit is much larger than the others, so gradients no longer sum to zero. The other, Softmax Cancellation, produces non-zero gradients even when student and teacher logits are equal. The authors present simple implementations that fix both bugs and report effects in several training settings.
Jonathan T. Barron, Dan Gnanapragasam, Michael Tschannen
In briefUniVOE: A Removal-Driven Framework for Interaction-Aware Video Object Editing
SummaryUniVOE is a framework for training both video object removal and injection models. An effect-aware remover, built by adapting video inpainting priors, converts real videos into object-absent versions to create paired data for training an injection model. A reward-based post-training step further aligns injection results with human preferences. Experiments report state-of-the-art performance on removal and injection, especially in cases with complex interaction effects.
In briefWorldForge V2: Aligning Camera Trajectories in Video Diffusion Models
SummaryThis paper presents WorldForge V2, a method that lets pretrained video diffusion models generate video along a target camera path while preserving scene appearance and object motion. It combines training-free guidance using warped references, optical-flow-based selection of motion-related latent channels, a dual-path artifact correction, v-cache to reduce cost, and Trajectory LoRA trained in two stages. The authors report state-of-the-art generation quality and trajectory accuracy for camera-controlled video generation across multiple datasets, with reliable control from very little training data.
In briefCausal Chain of Thought: Isolating Context to Curb Attention Leakage
SummaryThis paper introduces CausalCoT, a framework that splits a problem into a graph of atomic sub-questions and evaluates each one using only its direct causal parents. The goal is to make reasoning steps genuinely drive the final answer rather than serving as post-hoc justification. Across reasoning benchmarks, it keeps competitive accuracy while improving the causal sensitivity of intermediate steps over the strongest baseline.
Gabriele Dominici, Teague McMillan, Fazl Barez et al.
In briefCan an AI agent find flaws in a web text dataset? Auditing Slovak FineWeb2
SummaryThe paper uses a language model as a corpus debugger: a tool-using agent explores the Slovak portion of FineWeb2 through a read-only, budgeted API and turns each finding into an executable test that a native speaker validates. Over three runs, the agent found seven findings across four phenomena in 26.5M documents. The largest was 1,408,437 machine-generated documents (5.32%) that passed the pipeline's filters. The compiled test suite scans the full corpus in about 0.7 seconds without a GPU.
Marek \v{S}uppa, Ivan Vykopal, Jaroslav Kop\v{c}an
In briefHiP3R: Hierarchical Pose Graphs for Accurate Long-Video 3D Reconstruction
SummaryFeed-forward models can estimate camera motion and 3D geometry from short video clips, but long videos must be split into overlapping chunks, which creates inconsistent coordinate frames and duplicate camera estimates. HiP3R addresses this with a hierarchical pose graph: it first optimizes a chunk-level Sim(3) graph to set the global layout and scale, then fixes that layout and estimates one shared SE(3) camera per frame to merge overlapping predictions into a single trajectory. Experiments on multiple benchmarks show improved camera trajectories and 3D reconstruction. The method works with both chunk-based and streaming backends without additional training.
In briefBeNice!: Closing the Inference-Memory Gap in Looped Models via KV-Sharing
SummaryLooped architectures reuse the same layers repeatedly to improve quality without adding parameters, but each recurrence normally creates its own KV-cache, so memory grows with depth. This paper studies how to make looped models KV-shareable, using stochastic depth during pretraining and a post-training method, BeNice!, that trains the first recurrence to work with the final recurrence's KV-cache. The authors report that KV-sharing composes with other KV-compression methods and, in simulated low-memory long-context inference, improves cache hit rates, throughput, and p90 time-to-first-token.
Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick et al.
In briefQuestioning Instrumental Convergence in RLVR: OLMo-3 Checkpoint Study
SummaryThis paper examines whether reinforcement learning with verifiable rewards (RLVR) makes language models more likely to show instrumental-convergence behavior, such as self-preservation or resource acquisition. Using InstrumentalEval, the authors evaluate OLMo-3-7B-Think across SFT, DPO, 13 intermediate RLVR checkpoints, and the final model. The instrumental rate rises from 25.0% at SFT to 46.1% near RLVR step 200, then falls to 11.8% at the final checkpoint, below the SFT baseline.
In briefF2G-Pose: Geometry-aware feature lifting for direct RGB-D category-level object pose estimation
SummaryThis paper addresses estimating the 3D rotation and translation of objects from a category without instance-specific CAD models or reference views, using RGB-D input. The authors propose F2G-Pose, a single-pass framework that predicts pose, metric size, and camera-frame completed shape from one RGB-D observation. It reports state-of-the-art accuracy on SOPE and ROPE and runs at 28 FPS for pose-only inference.
WhyLifting foundation features into point clouds for single-pass direct pose prediction is a substantive improvement over prior methods, with strong benchmark comparisons on SOPE and ROPE.
In briefBilingual language models capture cross-linguistic influence in second-language reading
SummaryMonolingual language model surprisal predicts reading times but does not capture how a reader's first language shapes second-language reading. Using eye-tracking data from readers with 12 first-language backgrounds reading English, the authors compare bilingual models trained on the readers' L1 with a monolingual English baseline and bilingual models trained on other L1s. Matching the reader's L1 improves prediction of L2 reading times, with the advantage concentrated in late-stage measures, modulated by script overlap, and strongest among intermediate-proficiency readers.
Abderrahmane Issam, Irene E. Winther, Stefan L. Frank
In briefHow Ground Truth Labels Affect Retrieval Scores: Evidence-Aware Evaluation of a Breast Lesion RAG System
SummaryThis paper proposes an evidence-aware evaluation protocol for a retrieval-augmented generation system used on breast lesion management literature. Using an 18-document corpus and 18-query development benchmark, it compares chunk sizes, overlaps, and retriever configurations. The selected hybrid retriever placed accepted evidence in the top three results for 17 of 18 queries, with Recall@3, R-Precision, MRR@3, and nDCG@3 each at 0.94. Expanding relevance annotations increased evidence spans without changing rankings. The authors note the small development set requires independent validation.
Miguel Mochizuki Silva, Samuel Ismael, Mateus C. Dias et al.
In briefCRANE: Post-hoc counterfactual retrieval for multimodal classifiers
SummaryThis paper introduces CRANE, a post-hoc, model-agnostic framework that generates multimodal counterfactual explanations by retrieving real training examples from the target class. Because the candidates are actual training instances, plausibility is ensured by construction. Across four heterogeneous datasets and three fusion strategies, Multimodal Consensus Retrieval (MC-R) performed best or jointly best on three datasets, while unimodal strategies remained competitive when one modality dominated the decision boundary.
In briefSimulated Control Evals: using a language model to stand in for the environment when testing whether AI safeguards hold
SummaryThe paper introduces simulated control evaluations (SCEs), which use a language model to simulate tool responses in place of a real environment to test whether safeguards stop an AI model from causing harm when it intentionally tries. On LinuxArena, SCEs predicted safety with a root mean squared error of 3.6 percentage points, compared with 10.1 for BashArena. The authors conclude that deployment-specific SCEs can be more predictive than evaluations run in separate control settings.
WhyUses LLM-simulated tool responses for control evals, validated against execution on LinuxArena with lower prediction error than BashArena.
In briefBeyond Best-of-k: Prior-Consistent DAPS for Ambiguous Nonlinear Inverse Problems
SummaryThis paper addresses nonlinear inverse problems such as phase retrieval, where several measurement-consistent solutions can only be told apart by the prior. It shows that existing diffusion-based samplers may select a measurement-consistent solution that the prior does not support, a failure often hidden by best-of-k evaluation. The authors propose Prior-Consistent DAPS (PC-DAPS), which keeps multiple candidates in one sampling run and prunes weaker ones by prior consistency without ground-truth access. The method reduces prior-inconsistent samples and improves nonlinear problem quality while preserving quality and diversity on linear problems.
Merve Kocabas, Benjamin A. Burns, Andreas Geiger et al.
In briefJoint prompt and model optimization for cost-efficient multi-agent LLM systems (UniMASO)
SummaryThis paper presents UniMASO, a framework that jointly optimizes prompts and model assignments for each role in LLM-based multi-agent systems. It evaluates candidate updates through local, downstream, and global feedback, using cost-aware Pareto search to keep useful quality-cost trade-offs, with topology-aware co-evolution and cross-level trajectory memory. Experiments on five benchmarks and five execution topologies show competitive task performance at substantially lower deployment cost than the strongest homogeneous configurations.
In briefVERA: A dashboard for assessing LLM deployments against the EU AI Act
SummaryThis paper presents VERA, an open-source, self-hosted dashboard that helps non-engineering reviewers evaluate large language model deployments. It scores the twelve COMPL-AI requirements that current benchmarks can measure and records the energy used by each evaluation. In a within-subject study, the average assessment time fell from 56 seconds with raw output files to 30 seconds with the dashboard, and an internal survey rated usefulness 4.36 out of 5.
Robin Qu\'eriaux, Solveig Bachellery, L\'ea D\'el\'eris et al.
