In briefTokenEconomizer: An External Controller That Cuts Token Costs for Coding Agents
SummaryThis paper observes that frontier coding agents incur high token costs, and that when evaluated over sequences of realistic coding tasks, they show recurring environment-dependent token costs. The authors introduce TokenEconomizer, an external controller that decides which operational micro-skills to inject, which persistently irrelevant tools to prune, and when to intervene in unproductive failure loops. Across Claude Code, Codex, and GitHub Copilot CLI, it reduces average token cost by 18.4% to 30.9% with comparable task success, and costs fall by 6.4% to 21.2% in tasks such as web development and security auditing.
In briefImproving OCR Accuracy by Verifying Visually Similar Characters with Multiple Classifiers
SummaryThis paper addresses OCR errors between visually similar characters such as 'O'/'0', 'Z'/'2', and 'S'/'5'. It presents the Character Verification Model (CVM), a post-processing framework that uses binary classifiers trained on confusable character pairs to verify and correct OCR outputs. CVM does not rely on dictionary validation, so it suits serial numbers and license plates. Experiments with Tesseract, EasyOCR, and TrOCR on ten fonts report accuracy gains of 3.95 to 14.81 percentage points, an 87.3% correction rate for these errors, and a 2.3% false-correction rate.
Nuzhat Khan, Ab Al-Hadi Ab Rahman, Mohamad Anuar Kamaruddin et al.
In briefA playbook of misinformation tactics that sharpens judgement without fostering cynicism
SummaryThe study maps persuasion strategies recurring across 214,646 fact-checked claims in 51 languages into a misinformation playbook. Briefed with it, a language model and people in a pre-registered experiment became better at distinguishing unfamiliar claims, while their belief in true claims was unchanged. Averaged over all claims, however, the improvement was not statistically significant.
WhyProposes defending against persuasion strategies rather than facts, mapped from 214,646 claims; tested with a language model and a preregistered experiment showing gains for unfamiliar claims only.
In briefLEMON: Zero-shot monitoring of LLM misbehavior via equivariant learning on attention maps
SummaryThe paper introduces LEMON, an online graph neural network that reads attention maps from large language models and outputs a score after each token or window of tokens. Its predictions are invariant to algebraic symmetries between models, so it can transfer to new LLMs without re-annotation. Across 16 LLMs and six datasets, LEMON outperforms all baselines on hallucination and jailbreak monitoring in the zero-shot cross-model setting, improving by up to 14 ROC-AUC points.
WhyApplies geometric deep learning symmetries to cross-model attention monitoring, validated across 16 LLMs and six datasets with large gains over baselines.
In briefCoLU: A layer-selective method for making large language models forget specific information
SummaryThis paper proposes CoLU, a method for making large language models forget certain information while retaining others, without retraining from scratch. CoLU updates only one selected transformer layer and combines confidence-weighted forgetting with retain regularization. The authors also introduce Brief Layer Screening (BLS) to pick layers more cheaply, which they report reduces CoLU's runtime by 60.2-79.7%. On TOFU and MUSE benchmarks, the authors report better forgetting and retention than several compared methods.
Jingkai Guo, Ziyi Zhao, Chaitali Chakrabarti et al.
In briefHilbertGen-3D: topology-aware 3D shape generation using Hilbert-multifractal conditioning
SummaryThis paper presents HilbertGen-3D, a diffusion-based framework for generating 3D point clouds that adds global structural guidance through Hilbert-Multifractal Conditioning. The method computes multi-scale voxel measures, orders them along a 3D Hilbert curve, and estimates generalized fractal dimensions. On ShapeNet's Chair, Airplane, and Car categories, it reduces the 1-NNA deviation from 50% by up to 8.90 points and raises coverage by up to 5.88 points compared with TopoDiT-3D.
In briefARBOR: Finding Reasoning Failures in Long Chains of Thought by Tracking Reasoning States
SummaryLarge reasoning models solve hard problems with long chains of thought, which can fail through overthinking, underthinking, or unsupported deductions. ARBOR represents each reasoning trajectory as a tree of reasoning states and attributes each failure to the relation between a step and the state before it. The authors refine the three failure modes into subtypes and find that the tree state before a step separates failure steps from at-risk steps where no failure occurred. The paper also discusses how ARBOR can work with existing inference-time and training-data mitigations.
Huihan Li, Siyuan Wang, Naga Vamsi Ramana Dinavahi et al.
In briefChairNav: Pre-training and Personalizing Wheelchair Navigation Across Robot Platforms for Long Sidewalk Routes
SummaryThis paper presents ChairNav, a framework that adapts a general visual navigation model into a personalized autonomous electric wheelchair system through three training stages: pre-training on cross-embodiment datasets, mid-training on a few wheelchair demonstrations, and post-training from user interventions. The system is deployed on a commercial wheelchair with an RGB-D camera and an NVIDIA Jetson module. Reported results show an 18% lateral ADE reduction from pre-training and an increase in mean distance between interventions from 29.2 m to 57.4 m after post-training, with the system driving autonomously for 93.4% of time on kilometer-scale urban sidewalk routes.
In briefRefusal and Risk: How LLMs Respond to Suicidal Ideation
SummaryThis paper examines how eight instruction-tuned large language models respond to disclosures of suicidal ideation and self-harm across three datasets, with mechanistic analyses on six open-weight models. At baseline, higher response risk stems mainly from inattention to the disclosure and from refusal to engage. A clinician-guided response protocol consistently lowered response-level risk while also reducing refusals. The authors argue models should refuse only when warranted and engage safely when users disclose suicidal ideation.
WhyEvaluates eight instruction-tuned LLMs on suicidal disclosures, with mechanistic analyses and ablations in six open-weight models linking refusal to response risk; moderately novel with solid evidence
Drishti Goel, Soorya Ram Shimgekar, Jeongah Lee et al.
In briefFirst Passage Attention: forgetting by path extrema instead of decay sums in attention
SummaryThis paper proposes First Passage Attention (FPA), which replaces the summed decay along key-query paths in forget-gated softmax attention with a penalty based on the deepest dip of a learned scalar field. The authors show FPA generalises the Forgetting Transformer (FoX) and that a constant-dimension head can realise any subset of surviving keys. In a 110M-parameter language model, FPA matches FoX perplexity and achieves perfect passkey retrieval at eight times the training length.
In briefSpectral limits of polynomial graph filters on the long-range Ising benchmark
SummaryThe paper uses the Long-Range Ising Model benchmark (LRIM), where labels depend on long-range spin coupling by construction. It proves that readouts based on polynomials in the adjacency matrix hit an error floor that no increase in degree removes. Filters that weigh offsets the adjacency merges and resolve coupling across the full lattice escape this floor, and a strided filter with about fifty coefficients achieves much lower test error than the best published network.
In briefNormalizing Flow Maps: few-step generative models with exact likelihoods
SummaryThis paper introduces Normalizing Flow Maps (NFMs), a class of exactly invertible pushforward maps that generalize normalizing flows and flow maps. NFMs learn the integrator of an underlying dynamical system and provide efficient exact likelihood evaluation by construction. Training uses two consistency axes, pointwise matching and distributional matching, to derive distillation objectives. NFMs achieve state-of-the-art sample quality among parameter-matched normalizing flows on CIFAR-10 and ImageNet-64 conditional generation, support likelihood-based RLHF, and outperform prior normalizing flows on Boltzmann Generator equilibrium sampling benchmarks.
Francesco M. Ruscio, Timothy James Hitge, Fred Zhangzhi Peng et al.
In briefGlobeBench: A Benchmark for Testing Text-Based World Models in Environment Simulation
SummaryThis paper introduces GlobeBench, a benchmark of 1,632 tasks across five domains and nineteen simulation skills for evaluating text-based world models that simulate reinforcement learning environment state transitions. Models must predict the environment's response to an action given the full agent trajectory, and are scored by comparing against execution-derived references using a hybrid of deterministic gates and a semantic judge. Among 20 models, GPT-6 Astra achieved the highest success rate at 49.57%, while accuracy fell to an average of 5.1% when changes had to propagate to dependent objects. The authors note that 631 tasks remain unsolved by any model.
Nicholas Saban, Charvi Bannur, Mariya Vasileva et al.
In briefBrain-aligned multimodal AI models rating social and emotional content in cartoon films
SummaryThis paper compares two types of multimodal large language models with human ratings of Theory of Mind and emotional valence across three short cartoon movies, examining their temporal agreement. The authors also use each rating source to predict regional fMRI brain responses. The models capture aspects of human temporal variation but show systematic distributional differences, with valence generally aligning more strongly than Theory of Mind.
In briefIdentifying query templates from data provenance footprints and structural descriptors
SummaryThis paper proposes identifying query templates using data provenance, aggregating tuple-level Lineage and Why-provenance across full query results into query-level profiles. It further introduces Provenance Structural Descriptors (PSD), a small set of numerical features summarising these profiles. Experiments on 12,500 query instances generated from parametrised TPC-DS templates show templates can be identified from provenance alone, with accuracy improving as more history becomes available. PSD reportedly achieve high classification accuracy with limited training history and low test-processing cost.
In briefNoveltyScope: A multi-agent tool that assesses scientific paper novelty using retrieved evidence
SummaryNoveltyScope is a prototype that compares a scientific paper with retrieved prior work and produces a structured report and a five-level novelty score. It searches a closed local corpus and uses four scoring roles that rely on the retrieved evidence. On 50 class-balanced RINoBench samples, the multi-agent configurations had lower mean absolute error than single-LLM configurations for both Gemini and Qwen3.
In briefDiscrimination and polarization arising among learning AI agents
SummaryThis paper theoretically studies how discrimination and polarization can emerge in a population of learning agents. Each agent maps issue embeddings to binary opinions and updates its trust in other agents through an online Bayesian rule. By adding group labels and a group bias parameter, the authors use statistical-mechanics order parameters to show four distinct collective states in simulations.
Nestor Caticha, Jo\~ao Henrique Sant'Ana, Felippe Alves et al.
In briefThinking from Above: Bird's-eye-view tokens for efficient spatial reasoning in video AI
SummaryExisting video multimodal models organize visual information mainly over time, which limits spatial reasoning for embodied agents. This paper introduces Cambrian-BEV, which uses depth and camera pose to back-project frame features into a shared bird's-eye-view grid and average features per cell. The method compresses 128 frames of roughly 100K patch features into 729 BEV tokens, cutting LLM prefill time by 2.9x and peak memory by 3.6x. It scores 63.48 on VSI-Bench, 8.5 points above a matched frame-token control while using 12.6x fewer visual tokens.
Faizan Siddiqui, Nirav Savaliya, Sahith Reddy Chada et al.
In briefRepair and Retry: Compiler Feedback in Formal Theorem Proving
SummaryThis paper studies whether compiler error feedback helps language models repair failed Lean proofs. Through controlled one-step experiments on difficult MiniF2F problems, it tests Qwen3.6-27B and DeepSeek-Prover-V2-7B. Showing a failed proof reduces success and increases copying, while adding the matching compiler error partly recovers performance on Qwen. No tested feedback condition outperforms simply resampling with a retry instruction.
In briefCost-Aware, Reliability-Constrained Personalized LLM Routing via Online Preference Learning
SummaryThis paper studies personalized LLM routing, where a router chooses among models with different costs and capabilities while user preferences must be learned during deployment. The authors formulate the task as a constrained contextual bandit and propose an algorithm that balances inference cost against response quality. Under a logistic preference model and a strict-feasibility condition, they prove sublinear expected cost regret relative to an optimal stationary randomized contextual oracle, plus a high-probability guarantee on realized cumulative satisfaction. The framework is presented as extendable to other online model-selection problems with reliability requirements.
In briefOptimal ReLU network approximation of smooth functions
SummaryThis paper studies uniform approximation of Hölder ball functions by fully connected ReLU networks. For every s>0 and D≥1, it constructs networks with width O(N), depth O(L), and parameter magnitude O(1) that achieve uniform error of order (N²L²logN)^(-s/D). The construction relies on new encoding and decoding mechanisms that use higher-order finite-difference bounds on local Taylor coefficients. As a statistical consequence, empirical risk minimizers over the corresponding clipped network classes attain the minimax regression rate n^(-2s/(2s+D)) without extra logarithmic factors.
WhyCloses the ReLU approximation efficiency gap for all s>0 and D≥1 with new encoding mechanisms, yielding optimal rates and a sharp minimax regression consequence.
In briefTesting tools in separate branches, then keeping only what works
SummaryLLM agents often must find the right tool among thousands, and when several candidates could fit, only running them reveals which is correct. This paper introduces BRIA (Branch-Inspect-Adopt), which investigates each candidate tool in an isolated branch, then has a task agent review the execution evidence and adopt only useful results. On MCP-Atlas, BRIA reaches an average claim score of 89.77 versus 75.40 for MCP-Zero, and it matches or exceeds the strongest retrieval baseline across all tested settings.
In briefMultimodal filesystem memory for long-horizon embodied agents, keeping visual details alongside searchable text records.
SummaryThe paper presents MFM, a memory architecture for agents that stores episodic, spatial, semantic, procedural and working memory, linked to original multimodal episodes through searchable text records. Agents retrieve past experiences via contextual cues and revisit raw observations to recover details. Tested on long-horizon Minecraft and computer-use tasks with three base models, MFM with GPT-6-Astra completes Mine Diamond from Scratch and reaches 16 of 24 milestones on Ender Dragon (best baseline 12.7), and scores a mean reward of 0.440 on CUABench-KiCad (best baseline 0.400; no persistent memory 0.267).
In briefAlphaBet: Reinforcement Learning for LLM Event Forecasting Agents
SummaryAlphaBet is a reinforcement learning framework that trains one language model to search for evidence, reason, and report probabilities at several dates before an event resolves. Each dated forecast is scored against the final outcome with a proper scoring rule, so one resolved event supervises decisions under different information states. On GLM-4.5-Air, the Brier score drops from 0.731 to 0.626 on Forecast-Dojo and from 0.612 to 0.528 on FutureX.
In briefAuction method for reassigning airport time slots without money
SummaryThis paper studies how to reallocate indivisible resources, such as air traffic slots, when capacity cuts make existing assignments infeasible and cash compensation is impractical. It proposes a sequential scheme using reusable artificial credits and an ascending-auction mechanism, and proves properties including individual rationality, nontrader neutrality, and Pareto efficiency. Simulations of airport slot reallocation report lower total modeled delay and cancellation costs than a ground-delay baseline.
In briefFlexUMM: Flexible Representation Learning for Unified Multimodal Models
SummaryUnified multimodal models combine understanding and generation in one backbone, but these two abilities can produce conflicting representations. This paper introduces FlexUMM, a framework that dynamically decides when to share computation, when to route specialized capacity, and when to stop processing. The authors report that this flexible allocation improves accuracy by reducing harmful competition between functions and improves efficiency by skipping unnecessary computation. It is also described as a model-agnostic plug-in that can be applied to existing unified multimodal models.
In briefEgoP²C: A single vision-language model that learns to plan, perceive, and compose for spatial reasoning in first-person videos
SummaryThis paper addresses answering complex spatial reasoning questions from egocentric video, where a model must decide what evidence to gather and how to combine it. EgoP²C trains planning and perception teacher models with Dual-LoRA, then distills their behavior into a mixture-of-experts student through segment-level multi-teacher on-policy distillation. The resulting student runs the full plan-perceive-compose workflow without external spatial tools. Across three benchmarks, students built on Qwen3-VL-8B and SenseNova-SI-1.3-8B outperform their base models by 9.2 and 7.5 percentage points on average, while cutting inference FLOPs by about 66% and running 1.29 to 1.49 times faster than the Dual-LoRA teacher.
In briefWhose Polarization? Auditing Fairness in Network Polarization Measures
SummaryThis paper audits seven families of network polarization measures for fairness to minority communities, using two criteria: measurement invariance and structural exogeneity. It introduces Network-Weighted Distance (NWD) to fill a gap in the measure space. No measure passes all applicable checks, and minorities are systematically scored as more polarized than equivalent majorities. The authors trace this bias to majority structure affecting opinions and to structural change affecting structural measures, and offer guidance for reporting polarization scores.
WhyProposes a fairness audit and NWD measure, testing seven polarization families on synthetic and real networks, revealing two channels that systematically overscore minorities.
Stanis{\l}aw St\k{e}pie\'n, Michalina Janik, Akrati Saxena et al.
In briefDeepfake Face Detection: Comparing Deep Learning Models and Visualizing Their Decisions
SummaryThis paper compares two deep-learning approaches for detecting whether a single face image is fake: a small convolutional model trained from scratch and a pretrained MobileNetV2 adapted by transfer learning. Experiments used a balanced dataset of 140,000 real and fake face images. Under the recorded validation protocol, the pretrained model reached 95.90% accuracy versus 84.15% for the scratch model, and Grad-CAM was used to visualize influential image regions. The reported figures are validation results, not held-out test estimates.
In briefGossip Agents: Reputation Spreading via Opinion Dynamics on Changing Networks
SummaryThis paper studies how self-interested learning agents can sustain cooperation by sharing reputational information about third parties, a process called gossip. Agents form private opinions from limited observations and exchange them with neighbors using DeGroot opinion dynamics, on a communication network that changes as agents interact and move, with a tunable broadcast radius. The framework is tested in a Prisoner's Dilemma and in the Harvest environment. Results show that how far reputation information reaches shapes cooperation, and in Harvest, reputation diffusion increases environmental rewards and leads to more peaceful, sustainable behavior.
In briefNot All Graphs Are Maps: A Framework for Learning App Navigation on Phones
SummaryThis paper presents TAG-Nav, a framework that adds reusable destination memory to existing mobile GUI agents. Before any task is given, it explores an app offline and stores app functions, target pages, and routes in compact records, then uses them to guide navigation for future tasks without task-specific demonstrations. The authors also introduce DestBench, a navigation-only benchmark, and report that graph memory raises the strict navigation success rates of UI-TARS-1.5-7B and Qwen-VL-Max.
Adamo Orsini, Seyed Mehdi Ayyoubzadeh, Sadman Rakib Pinon et al.
In briefMemoProAgent: Memory-Driven Proactive Assistance for Timely Interventions
SummaryThe paper studies how assistants should use long-term personal memory to decide whether and when to proactively speak, rather than only answering queries. The authors introduce MemoProBench, a benchmark pairing egocentric video with year-long, distractor-rich memories and scoring both content and timing. Existing baselines perform poorly, at most 15%, while MemoProAgent raises strict accuracy to about 37.5–38% without training and 42.4% after reinforcement learning post-training.
In briefFrom Relevant Memories to Sufficient Evidence: Rethinking How LLM Agents Use Memory
SummaryThe paper observes that retrieving relevant memories does not always provide enough evidence to answer a query. It proposes QEMem, a framework that builds query-specific evidence sets through discovery across memory granularities, structured organization with source references, and sufficiency-guided refinement. On the LoCoMo benchmark with Claude Opus 4.5, QEMem reaches 91.43% LLM-judge accuracy, outperforming the strongest evaluated baseline, MemR³, by 7.6 percentage points. The authors also report improved evidence coverage and strong results on questions requiring cross-session evidence integration.
In briefDiverse Reasoning Paths Matter: Symbolic Equivalence Augmentation for RL
SummaryThis paper examines whether the diversity of chain-of-thought reasoning steps in supervised fine-tuning data affects the performance of later reinforcement learning with verifiable rewards. The authors introduce EquivAug, a data augmentation framework that uses symbolic equivalence transformations to create varied yet provably correct reasoning paths. Experiments on GSM8K and FOLIO across model scales report that EquivAug yields the highest path diversity after SFT and the best accuracy after GRPO, and the authors conclude that SFT chain-of-thought diversity is positively correlated with downstream RLVR performance.
In briefFinding Which Human-AI Conversations Need Costly Expert Correction
SummaryThis paper studies how to identify model responses in human-AI interactions that need correction. An expensive "editor", such as a human expert or a stronger AI, can review and modify responses, so the key question is which interactions to send to it before knowing whether an edit is needed. The authors formalize this as Edit Discovery Optimization (EDO) and propose EDIT, an online algorithm that respects an editor budget under partial feedback. On four reasoning and factual question-answering benchmarks, EDIT reports a better trade-off between editor cost and both error types than direct LLM-based batch selection.
Alireza Mirrokni, Shayan Kiyani, George Pappas et al.
In briefSecondhand Stress: a study of how reading others' distress affects users in graduate admissions forums.
SummaryThis paper analyzes 171,981 posts and 1,130,413 comments from two Reddit graduate admissions communities across three cycles, drawing on stress crossover theory. Using propensity-score stratification and baseline-adjusted comparisons, it finds that users who engaged with distress disclosures during application preparation later showed higher classifier-estimated anxiety, depression, and stress expression, along with more commenting and lexical redundancy. No significant associations were detected among downstream users who interacted with previously exposed users.
In briefIs a Linear Probe Evidence of a Linear Representation? A paper questioning whether linear classifiers can show that vision models encode generative factors linearly.
SummaryThe paper argues that high linear probe accuracy does not establish linear encoding of generative factors, since sufficiently rich feature maps, even random ones, permit linear readout. The authors propose three diagnostics (reverse predictivity, spectral concentration, and out-of-distribution generalization), validate them in a simulation with known ground truth, and apply them to six pretrained vision transformers and other datasets. Pretraining yields structure beyond a random kernel but falls well short of a true linear world model.
Nadir Kutluozen, Christian Intern\`o, David Klindt
In briefA staged, anatomy-aware method for segmenting intracranial aneurysms by vessel in CT and MR angiography
SummaryThis paper presents a six-stage pipeline for the TopAneu 2026 challenge, which segments intracranial aneurysms at the voxel level on CT and MR angiography and assigns each to one of over fifty vessel classes. The stages cover modality routing, vessel segmentation, candidate localisation, lesion segmentation, instance formation, and per-instance anatomical classification. The authors report stage-wise performance and an ablation study on vessel priors and class-aware sampling.
Gniewosz Drwiega, Wojciech Szymanski, Marek Wodzinski
In briefEvidence-gated bilingual RAG for consumer banking search that answers, flags mismatches, or declines
SummaryThis paper presents a bilingual retrieval-augmented generation system for searching consumer banking content in Arabic and English, running on local infrastructure. It combines web-content refinement, hybrid retrieval, intent-aware evidence selection, reranking, and a three-way Answer, Related Mismatch, or No Answer decision. On 210 labeled queries, Top-1 hit rate rose from 50.48% to 79.05% over a hybrid baseline. The authors state these results are preliminary, since queries paraphrase a small reference set and the end-to-end test is small and judged automatically.
Ahmed Mohamed Mahmoud Rayyan, Rana Mahmoud Mohammed, Aya Abdallah Abdelmonem et al.
In briefCost of Switching Optimizers Between LLM Training Stages
SummaryThis paper examines whether switching optimizers between pretraining and later training stages hurts LLM performance, focusing on AdamW and Muon. The authors pretrain 1.2B-parameter Llama-3 models from scratch while varying the share of updates from each optimizer, and interpolate between them during continued pretraining and supervised fine-tuning. They find little evidence of a mismatch penalty, with final performance largely explained by the independent effects of each stage's optimizer, and downstream results generally improving as Muon's share during fine-tuning increases.
Xinyu Zhao, Davis Wertheimer, Ashutosh Sharma et al.
In briefLGGS: Faster Brain MRI Reconstruction Using Prior Scans as Guides
SummaryThis paper proposes LGGS, a deep-learning method that speeds up MRI image reconstruction by using a patient's earlier scans as guidance. It uses learned phase-invariant circulant attention to match corresponding features across exams, making it robust to misregistration and phase variation. On longitudinal T1-weighted multi-coil brain MRI data, LGGS outperformed the evaluated unguided and longitudinally informed methods across acceleration rates, while using substantially fewer learned parameters.
Nikola Janju\v{s}evi\'{c}, Hersh Chandarana, Yao Wang et al.
In briefSAFETYEDIT: A Sparse-Target Method for Jailbreak Backdoors in LLMs
SummaryThis paper studies model-editing backdoors that induce harmful responses under a trigger while keeping normal behavior on untriggered inputs. The authors propose SAFETYEDIT, which optimizes a single refusal-related MLP coordinate and uses KV-cache marking in early attention layers to carry a prefix trigger signal to later instruction positions. Across four open-weight LLMs and three dataset groups, it reports the highest triggered Guard ASR in 11 of 12 comparisons and the highest mean StrongREJECT score in all 12. The authors also report that ablations and behavioral audits show generally minimal changes in untriggered behavior and general capabilities in the tested configurations.
In briefLearning concept embeddings for new concepts from cross-linguistic colexification patterns and short definitions
SummaryThis paper presents an inductive framework for learning concept embeddings. It combines the structure of full and partial cross-linguistic colexification networks with semantic information from short concept descriptions. This allows unseen concepts to be embedded based on a short definition. The resulting embeddings correlate with various cross-linguistic similarity or relatedness measures and generalize well to concepts absent from the training data.
In briefHard Instances on Demand: Using Optimization to Generate Difficult Test Cases for Algorithm Experiments
SummaryThe paper argues that conclusions from algorithm experiments depend on the test instances used, and that random instances can be misleading. The authors treat instance generation as an optimization problem, using solver running time as the objective and a black-box optimizer (Nevergrad) to search for hard instances. On four NP-hard problems (satisfiability, traveling salesperson, Hamiltonian cycle, and non-permutation flow-shop scheduling), the evolved instances were harder than the hardest of a budget-matched random draw.
WhyFrames hard-instance generation as black-box optimization of solver runtime, with comparisons against random draws and curated benchmarks across four NP-hard problems.
In briefMeasuring Mechanistic Uncertainty in LLMs via Causal Robustness
SummaryThis paper introduces mechanistic uncertainty, which measures how stable the internal mechanisms behind a language model's answer are. It defines a causal robustness margin as the smallest internal intervention needed to change the output, implemented with Jacobian Lens concept steering vectors. Across three open-source models and five question-answering datasets, incorrect and uncertain answers showed smaller margins, and the measure complements output-based uncertainty methods.
Aneesh Komanduri, Majid Afshar, Timothy A. Miller et al.
In briefChecking whether ICU sepsis prediction models flag the same patients after different imputation methods
SummaryThe paper introduces alert-list concordance, which measures whether two data-preprocessing (imputation) pipelines flag the same high-risk patients, using Jaccard similarity of their top-k% alert sets. Tested on the PhysioNet/CinC 2019 ICU dataset (40,336 patients, 71.9% missingness) with nine imputation strategies and a native-missingness baseline. Many method pairs with no significant AUROC difference had alert-list overlaps of only 0.518 to 0.750, showing similar aggregate discrimination can accompany different patient selections.
Priyanka Siddappa, Muntaqim Ahmed Raju, Shifat Haider Al Amin et al.
In briefToMStories: Testing Language Models' Theory-of-Mind With Synthetic Stories
SummaryThe paper introduces ToMStories, a dataset of about 6,000 short synthetic narratives, each paired with a multiple-choice or open-ended question about a character's emotion, fear, motivation, or emotional wound. The psychological states are conveyed implicitly, explicitly, or through conflict between dialogue and description. Across thirteen open-weight models (1B–72B parameters), the largest models match humans on multiple-choice questions but not on open-ended ones. The authors argue synthetic data helps reveal how models interpret implicit states, but likely overestimates how well they interpret real humans.
In briefPlausible explanations can mislead AI judges of answers
SummaryMultimodal large language models increasingly act as judges, often assessing answers together with explanations. This paper introduces JUPE, a benchmark that keeps answers fixed while varying explanations, and tests 14 such judges. Ordinary plausible explanations mislead judges about as much as deliberately persuasive arguments, reducing rejection of incorrect answers by 12.7 and 13.8 percentage points for plain and chain-of-thought rationales. The authors call this failure plausibility capture.
In briefChoosing models to estimate and forecast freeway speeds across a network using few fixed sensors
SummaryThis paper compares methods for estimating current and next-hour speeds on all freeway links when only some links have fixed detectors, trained on historical probe data. The test network is the Dallas–Fort Worth freeway system, with 643 of 4,127 directional links carrying detectors. Eighteen methods from five families were evaluated under one protocol. Graph neural networks passing information along the road graph reduced current-hour error on unsensored links by about 17% relative to the historical average, and adding interpolation features narrowed differences between architectures to about 0.1 mph.
In briefStability of outer Nesterov momentum on stochastic pseudo-gradients
SummaryThe paper studies how outer Nesterov momentum, the number of local optimization steps, and averaging across workers affect the stability of training large neural networks distributed over many workers. For random quadratic problems it derives an exact mean-square stability criterion in any dimension, showing that the inner batch size decides whether more local steps stabilize or destabilize training. It also finds that the number of workers needed for stability can grow exponentially with local steps, and tests the findings on several networks.
Nishant Sharma, Laura Gomezjurado Gonzalez, Ganesh Talluri et al.
In briefNoise-shaped sampling and annealing for training binary and ternary AI models on edge devices
SummaryThe paper shows that the training loss of binary and ternary neural networks is structurally equivalent to the energy of a high-order Ising model, allowing Ising solvers to be applied to discrete-weight training. It proposes ΣΔ-Sampler, which embeds stochastic sampling and annealing as a noisy quantizer inside a first-order ΣΔ noise-shaping loop to keep quantization and sampling noise out of the low-frequency adaptation band. The authors report validation on several benchmarks, including CIFAR-10, ImageNet-1k, and MatMul-free language models up to 2.7B parameters. They state it reduces training memory by over 50% compared with standard quantization-aware training, making on-device training more practical.
In briefTrajectory-Guided Steering: Reducing Spurious Feature Bias in Clinical Text Classification
SummaryThis paper examines how large language models in clinical classification can rely on spurious features, such as gender terms unrelated to a diagnosis. It proposes Trajectory-Guided Steering (TGS), which uses how the model's answer preference changes across layers to decide, per document, whether and in which direction to correct a prediction, without counterfactual data. On 16 i2b2 clinical tasks, TGS fixes 17.7% of errors while breaking only 1.3% of correct predictions, and it reduces gender-swap inconsistency on the most affected diseases.
In briefWhen Does Uncertainty Measure RAG Hallucination Risk?
SummaryThis paper examines whether uncertainty scores can measure hallucination risk in retrieval-augmented generation. Tests on four instruction-tuned models across HotpotQA and SQuAD show that uncertainty is informative mainly when retrieved evidence constrains the answer. A combined uncertainty and semantic-drift score improves calibration.
In briefShift-Reliance Alignment: ranking vision-language models by the cues they trust
SummaryThis paper proposes Shift-Reliance Alignment (SRA), a method for choosing vision-language models (VLMs) for a target domain without running every candidate there. It measures each model's reliance on cues such as scene context by editing source images so the cue misleads, then combines this with how often the cue misleads on target data. SRA lowered the error of the selected model on Objects365 and OpenImages but not on COCO, and the authors release the CueDial benchmark.
WhyProposes a new cue-reliance and shift scoring method for selecting VLMs, validated on multiple datasets and pre-registered tests, though COCO results contradicted the prediction.
In briefProactGen: Learning proactive questions to clarify video generation intent
SummaryThis paper presents ProactGen, a framework in which a vision-language model is trained with reinforcement learning to ask clarifying questions based on an initial prompt and a video draft. The questions aim to elicit user intent that the original description leaves out, with rewards based on how well the final video matches a target reference. Experiments report at least 1.8 times the alignment gain of the strongest baselines over direct generation. A single round of ProactGen also surpasses the alignment of the best baseline after four refinement rounds, using 29% of its time.
In briefTree-species LiDAR recognition across different data sources
SummaryFOR-species20K is a large public LiDAR benchmark for classifying 33 tree species, combining 25 independently collected terrestrial, mobile and airborne datasets. The paper proposes a multi-resolution, multi-view ConvNeXtV2 model that raises official hidden-test accuracy from 79.5% to 85.0%. The authors show that when source datasets are held out, accuracy drops sharply, so the leaderboard score reflects in-distribution performance rather than generalization to unseen sources.
In briefBeyond Explicit Intent: Finding and Shaping Artistic Intent Through Human-AI Co-Creation
SummaryThis paper proposes a human-AI co-creation approach in which artists discover their own preferences by observing and selecting from a wide range of generated visual options. Their curatorial choices are then used in reinforcement learning, strengthening recognized directions so the model gradually adapts to individual aesthetics. Three user studies examine latent preference discovery, personalized alignment, and externalizing vague creative intentions.
In briefRarity of Singletons Matters: A Fragility-Aware Way to Estimate Unseen Species
SummaryThe paper addresses how classical missing-mass and unseen-species estimators, which depend on rare singleton observations, become unstable with small samples. The authors propose a fragility-aware framework that uses an entropy–KL optimization to adaptively regularize these estimators based on singleton fragility. Tested across 360 configurations with synthetic distributions and real language corpora, the method stays competitive in stable settings and improves accuracy and stability in fragmented, long-tailed, and small-sample cases.
Pragatheeswaran Vipulanandan, Kamal Premaratne, Dilip Sarkar
In briefAutonomous AI Discovery on the Busy Beaver Frontier of Computability
SummaryThis paper presents a Gemini-based agentic pipeline that automatically discovers accelerations of Turing machines, expressed as compact mathematical rules summarizing many execution steps, in the uncomputable Busy Beaver domain. The system combines pattern recognition, automated conjecturing, informal reasoning, and formal theorem proving in Lean. It autonomously discovered and formally verified novel Collatz-like accelerations for eleven machines in BB(6), which are provisionally classified as Cryptids.
Marco Dos Santos, Vivek Veeriah, Alex Havrilla et al.
In brief"Stuck on Shuffle": Whether vision-language models correctly match attributes to the right objects
SummaryThis paper studies attribute-object binding, which is whether multimodal models correctly link each attribute to its intended object, in retrieval tasks. The authors introduce SABER, a benchmark covering all four image-text modality pairings with scenes of up to ten objects, and find substantial binding failures in both CLIP-style and MLLM-based embedders. They finetune an MLLM-based embedder using only synthetic attribute-binding text, and the gains transfer to cross-modal and image-image retrieval. The model reportedly achieves state-of-the-art results on SABER and also improves on negation and instance-retrieval benchmarks.
Netanel Y. Tamir, Michael Green, Issar Tzachor et al.
In briefCan AI models forecast whether machine learning experiments will succeed?
SummaryThis paper tests whether frontier large language models can predict how well machine learning experiment ideas they propose will perform before training is run. The authors built a dataset of 288 LLM-proposed ideas across eight MLE-bench tasks and compared five models' predictions with actual post-training outcomes. Predictions were much better than chance but inaccurate, with the best model explaining only about half the variance (R² = 0.49), and all models tended to underestimate the size of changes.
In briefWeakly supervised discovery of motion styles for combining robot styles with task commands
SummaryThe paper presents CODES, a framework that discovers a repertoire of motion styles from diverse recordings and composes them with independent task commands for robot control. Recordings are retargeted to the robot, then weakly supervised consensus encoding and density-based clustering identify styles without a preset count, and reinforcement learning trains a composing policy. On a Unitree Go2 quadruped, the authors report improved kinematic coherence over baselines in two simulators and on physical hardware.
Heedo Woo, Antonio Marino, Eduardo Sebasti\'an et al.
In briefDiagnosing GNN failure modes using layer-wise probe entropy
SummaryThe paper proposes a post-hoc diagnostic that freezes a trained graph neural network, fits a linear probe to each layer, and tracks probe accuracy and normalized predictive entropy. Across real and synthetic graph settings, layer-wise probe entropy distinguishes beneficial, neutral, and harmful smoothing. Comparing shallow and deep models, the authors show that where entropy degradation begins separates structural mismatch under heterophily from oversmoothing caused by excessive depth.
Shanzi Bao, Mar Gonz\`alez I Catal\`a, Pietro Li\`o
In briefThe Model Knows Better: Defending Medical AI Teams Against Hidden Injected Instructions
SummaryThis paper studies prompt injection in medical multi-agent systems, where LLM agents discuss a patient case, and finds a single injected passage succeeds in up to 61% of cases. Collaboration does not consistently prevent these attacks, and existing defenses only partly work. The authors propose a training-free defense that separates patient findings from clinical judgments and checks each judgment against the findings alone. It lowers attack success on MedQA to 4-7% while maintaining accuracy on cases without an attack.
In briefFlat Loss, Evolving Features: Grokking in Modular Addition
SummaryThis paper studies whether a neural network can keep learning while its training cross-entropy stays nearly flat at a positive level, using modular addition as a standard grokking testbed. ReLU networks keep 100% training accuracy while hidden features change and held-out accuracy rises above 90%. The authors prove this coexistence for one-hidden-layer networks and show experimentally that it also occurs in deeper networks and a one-block transformer.
WhyProves and experimentally shows feature learning under nearly flat training cross-entropy, via finite-window theorems and multiple architectures, offering a fresh view of grokking.
In briefSAG2: Embedding-free multi-hop retrieval using BM25 and JEV
SummaryThis paper presents SAG2, a multi-hop retrieval system that avoids vector embeddings and relies on a single SQLite database, BM25, and JEV-based contextual evidence decisions. Passages selected at query time supply terms for further retrieval, removing offline embedding computation and LLM-based entity extraction. On 1,000-question subsets of MuSiQue, 2WikiMultiHopQA, and HotpotQA, SAG2 achieves the highest Recall@10 among the compared methods.
In briefReducing Demographic Gaps in How Well Hate Speech Model Explanations Match Human Rationales
SummaryThis paper studies explanation fairness in hate speech detection, showing that high accuracy does not guarantee equally plausible explanations across demographic groups. The authors propose Plausibility-EFG, a metric that compares model token attributions with human-annotated rationales to measure the largest group gap in explanation plausibility. Evaluating BERT, RoBERTa, and an adversarially trained ADV-RoBERTa on HateXplain, they report that ADV-RoBERTa improves performance and explanation alignment while reducing cross-group disparities.
Suryansh Talukdar (1), Taha Sarwat (1), Jonathan Rusert (1) ((1) Purdue University Fort Wayne)
In briefRemoving Effects of Faulty Vehicle Contributions in Federated Vehicular Networks Using Operational Symptoms
SummaryConnected vehicles learn a driving model through federated learning, and faulty contributions often appear only as delayed on-road symptoms that name no specific vehicle or round. The paper proposes AT-ONE, which builds unlearning into the aggregation rule, using curvature from recent global increments to score all contributions in one backward pass. A two-tier correction suppresses suspected effects each round and eliminates confirmed ones after a one-round probe. On three driving datasets, it leaves 43% to 61% less residual harm than the strongest baseline in eight of nine settings while staying within 0.36 points of full retraining in utility.
In briefFederated framework that automatically evolves agent harnesses by sharing reflections (explainer: distributed AI agents share experience and improve their own working scaffolding)
SummaryThe paper proposes Federation over Harness (FoH), a framework in which distributed LLM agents share experience through a central agent. The server agent automatically evolves the client harness for task solving and self-reflection, while also evolving its own harness for aggregating reflections and maintaining shared knowledge. FoH needs no ground-truth labels or validation data, learning only from the agents' self-generated reflections. Experiments on multiple benchmarks report improvements of 53.57% over a single agent with a fixed harness and 29.76% over harness self-evolution.
In briefNoise-robust forced alignment for spontaneous speech, improving word boundaries
SummaryThe paper shows that forced aligners perform worse on spontaneous, noisy speech, where fillers, laughter and disfluencies distort word boundaries. The authors use WavLM features, dedicated phone units for fillers and non-lexical vocal events, and a discriminative projection trained on clean and noise-corrupted features, self-bootstrapped from 149 hours of audio without manual boundary labels. On conversational and disfluent datasets it achieves lower word-boundary errors than the strongest of ten baselines, and it has fewer catastrophic misalignments than Montreal Forced Aligner.
In briefAdapting a frozen robot policy to new hardware without demonstrations
SummaryThe paper addresses transferring the flow-based VLA policy π0 to new robot bodies, which is hard because it was trained on fixed embodiments. A lightweight embodiment classifier is trained on a few calibration tasks to tell successful action chunks on the source body from those on the target body, and its gradient guides each denoising step of the frozen π0 without demonstrations or parameter updates. On the LIBERO simulation benchmark with four modified controllers, average target-body success rose from 72% to 80% across 12 suite-controller pairs, with a 31% relative reduction in failures.
In briefStructure-aware spatial kernels for Gaussian process models of distributed port-Hamiltonian systems
SummaryThis paper proposes a structure-aware kernel design for Gaussian process distributed port-Hamiltonian systems, a data-driven modeling approach for distributed dynamical systems. The method removes irrelevant state coordinates and separates state and spatial dependencies using a tensor-product covariance. Experiments on a nonlinear string with periodic, smoothly varying, and semi-piecewise stiffness profiles show that this decoupling consistently improves reconstruction accuracy, and profile-matched spatial kernels give larger gains in force-field and trajectory reconstruction.
In briefAligning brain encoding models of vision with individual similarity judgments
SummaryThis paper examines whether brain encoding models, which predict brain responses to images, can be fine-tuned so their representations better match individual people's perceptual similarity judgments. Using judgments of 100 images from 8 participants in the Natural Scenes Dataset, the authors personalize a generic similarity model (DreamSim) for each participant. Fine-tuning with personalized similarity information produced models most aligned with human judgments and also reproduced intersubject correlations of those judgments.
Daniel Chong, Haomiao Chen, Keith W Jamison et al.
In briefFixing LLM-written formal specifications with deterministic error checks
SummaryLarge language models often turn natural-language reasoning problems into formal specifications for symbolic solvers, but these specifications can run without error while still being semantically wrong. The paper finds that purely LLM-based diagnosis often misses such errors and sometimes breaks correct specifications. It proposes DRAGO, which uses deterministic, formalism-specific diagnosis and keeps a rewrite only when the diagnosed error weight strictly decreases.
Giovanni Vaccarino, Emanuele La Malfa, Dehai Min et al.
In briefWhich Annotators Do Perspectivist Models Predict Best? Agreement, Sparsity, and Performance
SummaryThis paper examines how perspectivist NLP models, which predict how different people rate text, perform across annotators. The authors compile the largest collection to date of perspectivist datasets with ordinal labels and evaluate several algorithm classes on them. They find models predict better for annotators who agree more with others, while annotator response rate shows no statistically significant positive correlation with performance.
London Lowmanstone IV, Vaagishwaran Sivakumar, Anwesha Samaddar et al.
In briefAutomated Detection and Grading of Rodent Seizures from Video and EEG
SummaryThis study evaluates deep learning models that detect seizures and grade behavioral severity using paired video and EEG clips from 20 rats in a soman-induced epilepsy model. With session-exclusive evaluation, detection macro-F1 scores reached about 0.97, while five-class severity grading reached roughly 0.56 to 0.66. Subject-exclusive evaluation lowered five-class performance, to 0.475 for video and 0.404 for EEG.
In briefCoverage cliffs in certifying tail quantiles of policies from logged bandit data
SummaryThis paper studies when logged contextual-bandit data can certify an upper tail quantile for a policy selected from the same sample. The authors build a PAC-Bayes Bernstein confidence envelope and a procedure, TailCert, that returns an upper confidence bound on the quantile or an inconclusive result. They derive a matched finite-class certification boundary, and simulations illustrate search-complexity shifts, winner's-curse failure of naive data reuse, and the power cost of sample splitting.
In briefDeterministic Bounds on Accuracy and Stability in Metric-Space Regression and Classification
SummaryThe paper introduces a geometric framework that quantifies ambiguity between inputs and targets using an average kernel size computed from data-dependent feasible sets. It derives computable, deterministic, non-asymptotic bounds on the best achievable empirical error and accuracy for a finite dataset, independent of the predictive model. Experiments cover tabular regression, image regression, and image classification datasets.
Nina Maria Gottschling, Jakob Gawlikowski, Michele Caprio
In briefMicro-Localization: Microscope-Based Mobile Robot Positioning from Ground Texture
SummaryThis paper shows that a mobile robot can locate itself with tens-of-micrometers precision by viewing natural ground texture through microscope optics. The authors relate magnification and image uncertainty to physical precision and build a system that handles calibration, camera modeling, and focus. Tested on a calibrated gantry across several surfaces, translation RMSE was 28 to 73 micrometers, and the system also ran on a low-cost robot and detected micrometer-scale thermal expansion displacement.
WhyAchieves tens-of-micrometer localization from microscopic ground texture, linking magnification to precision; validated across 1,007 queries per surface on a calibrated gantry.
Aaron Wilhelm, Marina de Albuquerque Marcon, Emith Uyanwatte et al.
In briefFRACTAL: a self-similar multi-agent architecture with formal bounds on communication, convergence, and depth
SummaryThis paper presents FRACTAL, a multi-agent architecture for recursive goal decomposition in which the same five-component organizational pattern repeats at every level of the decomposition tree. It gives formal results on communication cost under a two-phase protocol, a supermajority convergence guarantee, a depth bound, and a Bayesian escalation mechanism between layers. The authors do not report quantitative performance, offering only a qualitative worked example and discussing the model's limitations.
In briefWhen plans change, remember: replanning signals guide what robot AI keeps in memory
SummaryThis paper proposes RePlanMem-VLA, a method that lets vision-language-action robot policies decide what visual information to store by using changes in their own action plans as a trigger. Large disagreements between overlapping action predictions are treated as signs of task-relevant events, and a bounded memory keeps keyframes while filtering redundant frames. The authors report detector ROC-AUC values of 0.981 to 0.997 on simulated tasks with pretrained policies, and say task-relevant observations were retained until the later decision point in 100% of evaluated episodes. Real-robot replay data are also reported to support event selection and redundancy suppression beyond simulation.
Victoria Smelova, Daniil Kazachkov, Georgiy Kazantsev et al.
In briefHow generative AI produces and spreads beauty bias
SummaryThe paper audits 27 text-to-image generators released from 2022 to 2026, using prompts that never mention attractiveness, and compares 51,300 generated faces with 160,196 photographs of real people. Nearly all generators produced faces rated more attractive than real photographs by automated measures and by human raters, and groups associated with darker skin tones scored lower. In a preregistered hiring simulation (N = 570), exposure to high-attractiveness AI faces lowered hireability and trustworthiness ratings of real candidates.
WhyAudits 27 generators with 550 raters and a preregistered hiring experiment, showing AI faces shift real candidates' hireability ratings.
AJ Alvero, Deepak Varuvel Dennison, Alexandra Crespin et al.
In briefCFC26: An Evaluation Benchmark for Deploying Sonar-Based Fish Counting
SummaryThis paper introduces CFC26, a new dataset and evaluation protocol for testing automated fish counting from sonar video on rivers seen and unseen during training. It expands CFC22 to eleven sites across nine Pacific salmon river systems, from Alaska to Northern California. The authors also propose two counting metrics, nMNE and nMANE, which reveal a previously hidden directional bias between upstream and downstream counts. The dataset and code are publicly available.
Madison Van Horn, Suzanne Stathatos, Sevan Brodjian et al.
In briefCultural Competence: How well LLMs detect culture-specific stereotypes
SummaryThis paper systematically tests how effectively large language models detect stereotypes specific to particular cultures. It varies the prompting language, instruction-tuning language, and fine-tuning dataset size, using US, Dutch, and Chinese cultural contexts. The models do not reliably account for cultural context and tend to overpredict stereotypes. Non-English prompts improve cultural consideration, additional instruction tuning reduces false positives, and training data size affects detection non-monotonically.
Maximilian Splieth\"over, Jannes L\"otfering, Dominik Bachmann et al.
In briefSPOT: Speeding Up Diffusion Language Models via Predicting Token Commit Order
SummaryThe paper addresses the speed-quality trade-off in diffusion language models (DLMs) by introducing SPOT, a 0.62M-parameter drafter that predicts the order in which positions are committed and proposes a tree of candidate continuations verified in one target pass. It also introduces SPOT-Sets for dynamic threshold decoding, which drafts whole commit sets. On SDAR-8B, the authors report up to 6.8x fewer target passes and up to 3.4x higher throughput than standard decoding.
WhyProposes commit-order prediction for speculative decoding, with broad experiments across six diffusion LMs showing lossless speedups and strong novelty.
Eliezer Shlomi, Eliron Rahimi, Avi Mendelson et al.
In briefHeRE: A Hebrew Relation Extraction Benchmark Built by LLM-Guided Denoising of Knowledge Graph Alignments
SummaryThe paper notes that LLMs perform well on English relation extraction, but low-resource languages lack standardized benchmarks for assessing them. The authors propose a language-agnostic method that denoises noisy alignments between knowledge bases and texts using a natural language inference task, and they compare five denoising strategies. Using this pipeline, they release HeRE, a 1.5-million-instance silver dataset for Hebrew relation extraction.
Ron Keinan, Amir David Nissan Cohen, Reut Tsarfaty
In briefDon't Retrain, Remix: Using Mixture Hypernetworks to Adapt Web Agents to New Interfaces
SummaryThis paper introduces WebMix, a mixture hypernetwork that adapts web agents to unseen interfaces by composing previously trained low-rank adapters without gradient updates on the target. A scout agent explores the new interface and writes a compact manual, which a router uses to predict mixing coefficients over the adapter bank. Experiments with Qwen and Llama models on TimeWarp and WebArena-Lite-v2 show WebMix matches or outperforms in-context learning and LoRA baselines. On TimeWarp domain transfer, it gains about 8 percentage points on average over the best baseline.
In briefDo reasoning AI models handle everyday logic as well as formal logic?
SummaryThe paper introduces a benchmark based on linguistic and pragmatic reasoning phenomena to test how well large reasoning models handle everyday reasoning. It compares them with non-reasoning models under controlled prompting conditions. The authors report that reasoning models often perform worse when everyday reasoning departs from formal logic, and that some non-reasoning models with chain-of-thought prompting outperform them. They conclude that current reasoning models do not reliably capture everyday reasoning, pointing to limits in post-training.
Kentaro Ozeki, Risako Ando, Takanobu Morishita et al.
In briefFlashControl: A one-step controllable image generator built from a simpler single-stream teacher model
SummaryThis paper addresses slow diffusion-model inference and the difficulty of distilling controllable generation models that use dual branches. Instead of adding extra conditional branches, the authors finetune a subset of layers in a text-to-image diffusion model, producing DeltaControl, a single-flow multi-step model that the paper says matches dual-branch performance. Applying step distillation to this teacher yields FlashControl, which performs spatially controllable generation in a single forward pass.
In briefPearl-Chat: A Luganda Language Model Built From Scratch in JAX and Flax NNX
SummaryThis paper presents Pearl-Chat, an approximately 8.5M-parameter causal decoder-only transformer designed for Luganda, a low-resource Bantu language spoken in East Africa. The authors implement the model in JAX and Flax NNX and train a dedicated 8,192-token byte-level BPE tokenizer, which achieves 1.60 tokens per word, about half the token count of LLaMA-3 and Gemma. The model is trained in two stages, pre-training and instruction fine-tuning, and reports high ROUGE-L and BLEU scores on a held-out benchmark. Weights, tokenizers, and code are released as open source.
In briefLiteral Lock-In: Test Language Effects on Bengali Idiom Scorers
SummaryThis paper audits non-generative decision models that output probabilities over given options, testing their tendency to choose literal readings of Bengali idioms. Adding an idiom's own literal gloss to the options drew the models toward it, accounting for 65% to 96% of each model's remaining Bengali errors. On the same paired items, the models were 1.9 to 6.4 times more drawn to literal readings in Bengali than in English.
In briefContext-aware English-to-dialectal Arabic dialogue translation using QLoRA-tuned NileChat-3B (AlexandriaX-2026 Subtask 1)
SummaryThis paper describes team Alkhder's system for Subtask 1 of the AlexandriaX-2026 shared task, which translates English dialogue turns into thirteen Arabic dialects. The team fine-tuned NileChat-3B-Base with QLoRA on the Alexandria parallel dialogue corpus, using a single set of adapter weights for all dialects. Each turn is conditioned on target country and dialect, domain, roles, gender direction, and preceding context. Under the constrained track, the system reaches a country-macro-averaged spBLEU of 20.36 and chrF++ of 36.64.
In briefLearning shared structure for cryo-EM inverse problems
SummaryThis paper proposes learning a common representation from cryo-EM particle images using self-distillation, so that pose estimation, heterogeneity analysis, and reconstruction can share one model. The authors introduce the EMber vision transformer family, pre-trained on over 50 million experimental particle images from 214 EMPIAR datasets plus simulated projections from 268,000 molecular assemblies. The paper reports gains of 14.85 and 14.26 percentage points on conformational and compositional discrimination, with reference-quality reconstructions reached in fewer epochs.
Nithin Chalapathi, Theo Sternlieb, Ethan Han et al.
In briefLearning smooth dynamics from positions and velocities with Kinematic Interpolants (paper)
SummaryThis paper introduces Kinematic Interpolants (KI), a framework that extends stochastic interpolants to infer nonlinear dynamics from snapshots in phase space that include both position and velocity measurements. The authors report that across several scientific datasets, KI reconstructs the underlying dynamics better than existing baselines. They also state that it needs only a fraction of the training time.
Yiheng (Roger) Jiang, Youssef Marzouk, Avishek Joey Bose et al.
In briefLearning User Representations to Better Simulate Real Users
SummaryThe paper argues that user simulators struggle to reproduce diverse real-user behaviors, which weakens AI assistants on underrepresented users. It proposes natural-language user representations that condition a simulator, evaluated with a conversation reconstruction task that compares user behavior in original and simulated conversations. It also introduces hindsight distillation, which trains a representation generator called UserRepGen-4B. Across 700 conversations from seven datasets, UserRepGen-4B produces representations that enable more faithful simulation than GPT-5.5.
Shuhaib Mehri, Philippe Laban, Michel Galley et al.
In briefPricing Buyers Using Features of Differing Trustworthiness
SummaryThe paper proposes a source-aware framework for contextual dynamic pricing, distinguishing verified customer features from self-reported ones that buyers may manipulate. The SALE policy uses verified features to localize price exploration and combines both feature types during exploitation to correct strategic manipulation. The authors establish consistent sparse estimation and a sublinear expected regret bound in a high-dimensional semiparametric setting, supported by simulations and real-data experiments.
WhyProposes source-aware dynamic pricing addressing strategic manipulation, with sublinear regret and uniform estimation guarantees in high dimensions; impact mostly confined to pricing research.
In briefalkhder's ImageEval 2026 Entry: Constrained Selection and Rotation-Averaged Scoring for Multimodal Hallucination Detection
SummaryThis paper describes a submission to Subtask 1b of ImageEval 2026, which asks systems to pick the one visually valid statement among three candidates paired with an image, in English and Modern Standard Arabic. The authors use a pre-trained 7B vision-language model in a zero-shot setting, with no task-specific training, and apply Latin-square rotation of candidates to reduce answer-position bias. The system reports CI scores of 0.0510 on English and 0.0960 on MSA test sets, reducing CI relative to official 3B baselines on devtest by 83.7% and 80.4%.
In briefReconstructing Geographically Diverse Architecture with 3D Foundation Models
SummaryThis paper introduces ArchWorld, a benchmark of 3D architectural landmarks with broad geographic coverage and rich metadata. The authors use it to analyze where current 3D foundation models make errors. They find performance varies by region, only partly matching training data representation, and also reflects regional architectural features and background distractors. Based on these findings, they propose an intervention that improves performance while reducing geographic disparities.
In briefPredicting how useful privileged information is for on-policy self-distillation of language models
SummaryThe paper studies how the content and construction of privileged information (PI) affect on-policy self-distillation for large language models. It evaluates 16 PI configurations across three multi-turn agentic benchmarks and three student models. It finds that information-dense trajectories and PI from stronger generators tend to help more, and proposes KL-based metrics to assess PI before training.
Kaijian Zou, Emiliano Penaloza, MohammadHossein Rezaei et al.
In briefAuditKIT: A unified framework for evaluating language models and AI agents
SummaryThis paper presents AuditKIT, a source-available evaluation library that records benchmark accuracy, judge-based assessment, runtime performance, tool use, and retrieval quality in one result format. For agents, it separates parallel from sequential execution and distinguishes verified task outcomes from trajectory-level diagnostics, marking missing evidence as unscored rather than failed. Seven case studies show how answer extraction, runtime behavior, and metric choice affect evaluation results.
Shreeyans Arora, Utsav Avaiya, Zera Lyngkhoi et al.
