In briefReinforcement learning for unit commitment on the EPRI benchmark, diagnosing over-commitment and speeding up MILP
SummaryThis paper is the first study of reinforcement learning on the 51-generator, 72-hour EPRI AI-accelerating Unit Commitment benchmark. It finds that standalone agents commit too many units because of their own exploration, which makes them costlier than mixed-integer linear programming (MILP). The authors show that the learned policy can accelerate an exact solver, and they propose a reporting protocol because dispatch method and benchmark scope shift reported results.
In briefAuditing Tool Traces in Agentic LLMs for Hidden Label Shortcuts
SummaryThis paper studies whether agentic LLMs making scientific predictions rely on hidden shortcuts in serialized tool results rather than genuine reasoning. The authors propose a two-stage audit: Per-Tool TF-IDF Decomposition localizes label-predictive patterns to individual tool outputs, and Token-Class Redaction tests whether the frozen model depends on them. Applied to TCR-pMHC binding prediction, the audit finds the strongest signal in identifiers within one tool result, and redacting them drops performance to chance.
In briefNeural networks with belief-function layers using Yager's combination rule to express 'I don't know'
SummaryThe paper proposes a deep evidential network in which every layer output is a valid Dempster-Shafer mass function, combined by Yager's rule. The authors argue that Dempster's normaliser causes gradients to diverge under high source conflict, whereas Yager's rule avoids this. The output mass on the ignorance set, m(Θ), serves as an 'I do not know' signal and ranks human-disagreement inputs in visual and linguistic tasks.
In briefA residential construction image dataset organized by CSI divisions for deep learning-based object detection
SummaryThis paper introduces ResCD-CSI, a manually labeled image dataset organized by the Construction Specifications Institute's divisional structure. It contains high-resolution images from active residential construction sites across the United States. A total of 16,365 objects across 14 major CSI divisions were annotated with bounding boxes for object detection tasks.
In briefModaliff: a single token-based diffusion model that generates across text, image, and audio.
SummaryModaliff is a token-based diffusion model for converting flexibly among text, image, and audio. It learns from widely available bimodal pairs and keeps each modality in its native latent sequence, using per-token attributes and role-based positional embeddings. One denoiser can handle diverse generation tasks without task-specific architectural changes. The paper reports the best fidelity among any-to-any models on most tasks, with substantially fewer parameters and training data.
In briefSpeech recognition for Saudi and Gulf Arabic using Gemma 4's audio encoder without changing its weights
SummaryThis paper studies a 305M-parameter audio encoder shared by Gemma 4 and EmbeddingGemma 2 for Saudi and Gulf Arabic speech recognition, keeping all pretrained weights unchanged. A frozen system trains only a small CTC reader, while an adapted system adds low-rank adapters, a hybrid CTC/attention reader, and joint decoding with a sub-word language model. Compared with Cohere Transcribe Arabic, the frozen system is slower in word error rate but faster on CPU, and the adapted system performs better on several test conditions.
In briefTurning ICU Risk Predictions into Staffing Decisions Under Uncertainty
SummaryThis paper proposes an uncertainty-aware framework that connects short-term ICU patient risk predictions to nurse staffing plans without changing the underlying prediction model. It uses Conditional Value-at-Risk (CVaR) to assess extreme staffing shortage exposure and feeds these risk-calibrated demand signals into an hourly rolling-horizon scheduling policy. Evaluated on MIMIC-IV trajectories, the framework shows a better cost-shortage trade-off than deterministic and reactive baselines.
In briefStopLoss: Teaching Coding Agents to Predict Failure and Stop Early
SummaryThe paper introduces StopLoss, a small external monitor that learns to recognize failing coding-agent runs from completed trajectories and their verifier outcomes. It ends a run early when failure is likely, saving resources. Across tested settings, it costs one to two orders of magnitude less per check than general-purpose judges and lowers cost per retained solve relative to completing every task.
In briefFaster algorithm for near-equilibrium in symmetric two-player games
SummaryThe paper presents the first unconditional family of polynomial-time algorithms for approximate Nash equilibria in symmetric bimatrix games. For any fixed δ>0, it computes a symmetric (1-1/√2+δ)-approximate equilibrium. This improves on the earlier (1/3+δ) guarantee by Kontogiannis and Spirakis (2011).
WhyImproves the polynomial-time approximation guarantee for symmetric bimatrix games past the 1/3 barrier with an unconditional algorithm family.
Ashkan Soleymani, Patrick Jaillet, Gabriele Farina
In briefFracTS: Hierarchical Autoregressive Model for Time Series Generation
SummaryThis paper presents FracTS, a coarse-to-fine hierarchical autoregressive model for generating multivariate time series. Adapted from fractal generative modeling for images, it captures global trajectory structure at coarse scales and local dynamics at fine scales. Across five datasets and seven baselines, it outperforms the baselines on marginal distributions, predictive ML utility, and correlation preservation in the unconditional setting. Predictive ML utility improves on average by 100.0% for unconditional datasets relative to the best baseline.
WhyAdapts fractal generative modeling from images to 1D time series with a coarse-to-fine hierarchy; evaluated against seven baselines on five datasets.
In briefWhen Empathy Misses the Goal: A Benchmark for Measuring Goal Displacement in AI Advice
SummaryThis paper examines how conversational language models often respond with warmth and reassurance while failing to address the user's actual goal, a failure the authors call goal displacement. They introduce the GDB benchmark, with 230 scenarios across nine domains and 4,134 human-scored responses from 18 models. Results show displacement rates vary widely across models and are strongly negatively correlated with Chatbot Arena Elo, and the authors argue the construct is distinct from social sycophancy.
WhyProposes goal displacement and a 230-scenario, 4,134-response human-scored benchmark, distinguished from sycophancy with reasonable empirical support and evaluation value.
In briefATEST: Adaptive Resource-Aware Test-Time Scaling for Efficient LLM Reasoning
SummaryTest-time scaling improves large language model reasoning by sampling multiple answers, but fixed budgets ignore query difficulty and the changing energy cost of inference under GPU load. The paper presents ATEST, which jointly weighs the expected value of extra samples against their current GPU cost to decide when to stop. Its default online mode needs no offline calibration, monitoring live GPU conditions and learning per-step energy cost as it samples. A calibrated variant instead picks a per-query budget from offline models and is competitive under stable loads.
In briefTransformers learn two competing sorting methods; only one generalizes to longer sequences
SummaryThe paper shows that two-layer transformers trained to near-zero loss on sorting converge to one of two mechanistically different solutions, depending on initialization. One uses both attention layers and generalizes perfectly to longer sequences despite training only on a fixed length; the other fails to generalize. The authors attribute the gap to input types that are rare at training length but more common at longer lengths.
Khashayar Gatmiry, Nathan W. Henry, Chloe H. Su et al.
In briefReDT: A Transformer for Efficient High-Resolution 3D Reconstruction from Multiple Images
SummaryThe paper introduces ReDT, a feed-forward visual geometry transformer that scales multi-view 3D reconstruction to high-definition inputs. It maps each view to a fixed-size set of basis tokens, computes correspondences in this compact space, and projects results back to the original resolution for dense prediction. Experiments across datasets from standard definition to 2K show better reconstruction quality and efficiency than prior methods. On ETH3D (2K), ReDT reports a 16× speedup over VGGT-Ω along with lower peak GPU memory and chamfer distance.
In briefBud ContextCompact 2.0: a method that summarizes older agent turns so small models with small context windows can keep working
SummaryThe paper presents Bud ContextCompact 2.0, a context compaction method built into the Bud agent runtime that replaces older turns with a summary so long sessions can continue on small language models. On recall questions over recorded sessions, a 1.2B model's accuracy rose to 0.335 with compaction, versus 0.247 with truncation and 0.176 with full history. On 100 SWE-bench Verified tasks, compaction raised the resolve rate on long sessions from 41% to 53%, but the gain disappeared under a 45-minute session budget because each summary took a median of 161 seconds.
In briefWhen Interpretations Persist: How Stored Guesses Shape Later AI Assistant Behavior
SummaryThis paper studies whether AI assistants that keep interpretations of a person's gaze-based activity affect their later outputs. With retrieved events, their order, and the target input held fixed, free-form interpretations amplified influence from history in both directions, helpful when correct and harmful when wrong. An evidence-bounded interpretation style, which separates observation, inference, uncertainty, and temporal scope, reduced unsupported and overcertain content and lowered harmful influence from wrong sources while keeping most helpful effects. The same pattern appeared with a second model.
Mingcong Ding, Chenyan Zhang, Zhuojiang Cai et al.
In briefTraining world models with random partial masking of observations to improve planning in pixel-based control
SummaryThis paper addresses representation learning in model-based reinforcement learning (MBRL) for continuous control from pixel observations. The authors propose a stochastic gating mechanism that samples a spatial binary mask from a Beta–Bernoulli random field and applies it to raw observations during training, exposing the encoder to varying levels of visibility. Evaluated on pixel-based tasks from the DeepMind Control Suite using model predictive control with open-loop rollouts, the method is reported to improve planning performance and data efficiency.
In briefKPI-aware multimodal anomaly detection for 5G and Open RAN telemetry
SummaryThis paper presents KAC, an anomaly-detection framework for 5G and Open RAN network telemetry. It aligns KPI-level residuals with compact operator-style text summaries, turning them into a learnable feedback signal over full-KPI temporal data. At a fixed 2.5 percent false-positive rate on TelecomTS, KAC achieves the highest recall among the evaluated residual-based methods. It is running in pre-production shadow mode with Nokia's KPI pipeline, scoring telemetry windows.
In briefHow Far Can an Agent Push? Trust-Safety Limits in Live Language-Based Economic Games
SummaryThe paper describes an entry into the GLEE competition using one protected agent, Milton, alongside four agents that run randomized trials of policy changes. In negotiation, learned pricing improved adjusted percentile, while in a preregistered persuasion experiment, recommending every unit raised sales where the receiver's slack made obedience rational and lowered them where it did not. The authors also report a close at the fixed stop, where a final-day pool shift affected all five agents' bargaining ratings.
In briefLanguage-specific features in multilingual LLMs appear across nearly all layers, not just the first and last
SummaryEarlier work suggested that multilingual LLMs encode language-specific features mainly in their initial and final layers, with intermediate layers handling language-agnostic representations. This paper re-examines that claim through four experiments, including fine-grained analysis of Portuguese and Spanish and a large-scale language identification test covering over 200 languages, using Llama 3.2 1B. The authors find that language identity remains strongly recoverable throughout the model, and that ambiguous strings tend to be classified as closer to Spanish than Portuguese.
In briefFrom Seeing to Reporting: A Theory for Reliable Visual Set Generation in Vision-Language Models
SummaryThe paper argues that when vision-language models report what they see, such as counting or locating objects, the reports can contradict themselves, for example listing 173 matching records while stating counts that sum to 188. The authors propose the FSR theory of visual reporting and inference-time procedures that improve frozen models' reports without retraining. Tests on dense synthetic scenes and 600 ScreenSpot-Pro GUI-grounding items show reduced count error and miss rates. The protocols, stored responses and analysis code are released.
In briefSteered Not Stirred: Effective Reward Tilting via Surrogate Reward Mixing
SummaryThis paper addresses steering generative models at inference time when the true reward is unavailable or costly to evaluate. It proposes MARTINI, which learns to combine several cheaper surrogate rewards from a small set of oracle rankings or scores, for use with standard steering methods. The authors give theory on surrogate information limits and report better oracle-tilt fidelity in language generation, ImageNet alignment, and protein-ligand cofolding, without inference-time oracle queries.
Alec Sargood, Olivia Gallup, Georg Maierhofer et al.
In briefFormal Raising: Building Checkable Ethical Rules into AI Development
SummaryThis paper proposes Formal Raising (FR), which represents selected moral values as explicit, machine-checkable properties used during AI development, evaluation, and action selection. It presents a C-based reasoning framework that evaluates candidate actions against prioritized ethical properties, resolves conflicts by a given order, treats strict rules as non-defeasible, and reports underdetermination explicitly. Examples cover conflicts involving life, truthfulness, harm, privacy, autonomy, and strict rules.
In briefA framework for judging how ready construction datasets are for AI
SummaryThis paper proposes a Construction Dataset AI Readiness Framework with 30 attributes across four dimensions: metadata completeness, data quality and representativeness, governance and accessibility, and AI utility. The authors tested it on 21 publicly available construction datasets, finding overall scores from 0.720 to 0.964 with a mean of 0.879. Automated suggestions matched manually verified scores 90.5% of the time. The framework aims to support consistent assessment of heterogeneous construction datasets and to identify readiness gaps.</summary>
In briefSelf-distillation method (Gold-Conditioned Distillation, GCD) for supervised fine-tuning
SummaryThis paper notes that supervised fine-tuning treats expert demonstrations as one-hot targets, which can suppress other plausible continuations the model would assign probability to. The authors propose GCD, where one model acts as both teacher and student: the teacher sees the gold response in its prompt, while the student sees only the original prompt and is trained to match the teacher's next-token distributions via forward KL divergence. Experiments on Qwen and Llama models reportedly show GCD improves new-task accuracy over SFT and generally reduces degradation on general benchmarks.
In briefFrom Proposal to Execution: Modeling Interaction Semantics in Shared Autonomy
SummaryThis paper addresses human-AI systems where candidate actions are first evaluated by another decision-making entity before execution. It proposes a decision framework that explicitly represents interaction semantics, using a joint action language and an interpretation function that maps actions and context to semantic outcomes. The framework is instantiated in shared human-AI autonomy, covering patterns such as advisory and authorization, and is stated to remain compatible with standard MDP planning.
In briefGroundedVDR: Checking whether video research agents truly use video and web evidence
SummaryThis paper introduces GroundedVDR, a benchmark for video deep research that checks whether models actually use temporal video information and retrieved web evidence. Questions are filtered through human review and restricted-access probing to remove those answerable from parametric knowledge or a single source, and each carries annotations linking video segments to web evidence. Evaluation scores video grounding, web evidence grounding, and answer correctness separately. Across eight multimodal agent systems, the best reached 76.0% answer accuracy, but none recovered the complete required web evidence for more than 8% of questions.
In briefSlimFormer++: Placing Token Mixing Where It Matters for Efficient 3D Medical Segmentation
SummaryThis paper studies where computation is needed in 3D medical image segmentation networks, finding high representational redundancy within blocks and little contribution from most intermediate encoder token mixers. Based on this, it proposes SlimFormer++, which uses windowed Mamba mixing only in the first encoder stage, residual channel MLPs in deeper stages, and a bottleneck mechanism that adds sparse long-range context. Across eight public CT, MRI, and ultrasound datasets (1,809 cases), it reports the highest mean Dice (87.84%), lowest mean HD95 (6.50 mm), and the fewest parameters (5.98 million) among 12 baselines.
In briefPCXR-Copilot: A device-aware system for pediatric chest X-rays covering segmentation, device suppression, and edema cues
SummaryThis paper introduces PediaDeviceSeg-1K, a dataset of 1,000 anterior-posterior pediatric chest X-rays from cardiac ICU patients, annotated with common ICU devices. Building on it, the authors present PCXR-Copilot, which outputs lung and device masks, optional device-suppressed views, and edema-related visualization maps and scalar indices. The full pipeline has been integrated into the TWIST platform at Boston Children's Hospital and runs in real time.
In briefPRIMA: A Pose-Guided Adapter for Zero-Shot Recognition of Social Body Behaviors
SummaryThe paper proposes PRIMA, a lightweight adapter that uses human pose only as a training-time signal to inject body-structure knowledge into frozen vision-language models, so inference needs only RGB input. It separates a model-specific projection head from a shared core, so adapting to a new VLM requires retraining only the head. The authors train on a micro-action dataset and evaluate by zero-shot prompting, reporting improvements on the training domain, on other backbones, and on two unseen indoor datasets. The results are reported only within the tested models and indoor datasets.
Mingcong Ding, Babette B\"uhler, Tim F\"utterer et al.
In briefMultimodal-Conditioned Latent Diffusion for Brain MRI Forecasting in Alzheimer's Disease
SummaryThe paper proposes MC-LDM, a latent diffusion model that synthesizes patient-specific brain MRI forecasts. It uses MRI, tau-PET, CDR and APOE4 biomarkers during training, but only baseline MRI, APOE4 and metadata at inference. Trained on 1,355 ADNI visits from 852 subjects, it reduces average regional volumetric MAE by 18% over the second-best baseline on an independent test set of 664 visits.
Antonio Scardace, Silvia Ingala, Sebastiano Battiato et al.
In briefConsistent Handling of LLM Calls Across Java Frameworks: A Reference Implementation and Benchmark
SummaryThe paper defines a six-clause reliability contract for how Java applications should handle failing LLM providers, covering retries, waiting, truncated answers, and metrics. The authors tested Spring AI and LangChain4j against it using a scripted OpenAI-compatible provider across fifteen failure scenarios. They then built a shared implementation at the HTTP boundary that all three tested clients use, and it met the contract in 90 of 91 cases. The one remaining case is a policy trade-off between retrying a hung provider and waiting for a slow first token.
In briefNoise-rejuvenated sequential Monte Carlo for test-time scaling of text-to-image flow matching models
SummaryThis paper proposes a test-time scaling method for text-to-image flow matching models, called noise-rejuvenated sequential Monte Carlo. Because flow matching uses deterministic ODEs, standard SMC does not transfer directly, so the method changes each resampled particle's source noise and builds a new intermediate state from its current endpoint estimate. A look-ahead mechanism refines particles before weighting. Experiments on Stable Diffusion 3, FLUX.1-dev, and Z-Image show consistent gains in reward-guided generation, and ablations validate both components.
In briefP(Yes): Test-time search raises decision model accuracy up to 15x
SummaryThe paper adds test-time search to frozen decision models without any training. The search scores complete candidate answers by the model's probability of answering "Yes" that they are correct, and seeks answers that score higher. On 128 logic puzzles, Qwen3.8-27B's fully correct puzzles rose from 2 to 30 and its per-blank accuracy from 27.8% to 48.5%.
In briefSZED's multi-task fusion method for detecting hateful Arabic memes in the ArGuard 2026 shared task
SummaryThis paper describes SZED's submission to the ArGuard 2026 Track A shared task on Arabic hateful meme detection. The task covers binary classification (A1) and fine-grained multi-label tagging (A2). For A1, the authors fuse MARBERTv2 and CLIP ViT-L/14 embeddings, while A2 uses a multi-task label-attention architecture. The reported macro-F1 scores are 0.7597 for A1 and 0.3142 for A2, with analysis of reproducibility, per-class errors, and cascade calibration.
In briefA study of whether vision-language models use visual knowledge in text tasks, and how self-distillation helps
SummaryThe paper asks whether training a language model for visual question answering also makes its text representations visually structured. The authors build shape-counting tasks given either as images or as text coordinates, and find that vision-aligned models do no better than text-only models on the text version. Probing confirms their embeddings carry no more visual structure. A proposed objective that distills in-context visual embeddings into text embeddings improves text-version performance without task-specific supervision.
In briefNear-optimal last-iterate convergence of SGD in the smooth interpolation regime
SummaryThis paper studies the last iterate of stochastic gradient descent with constant stepsize on smooth convex losses when gradient noise at the optimum is small or zero. The authors bound the expected excess risk after T steps for any stepsize below 2/β. In the interpolation regime this yields a near-optimal Õ(1/T) rate, improving on a prior O(1/√T) result for stepsize 1/β.
WhyA new second-moment comparison technique over logarithmic time scales yields near-optimal last-iterate SGD rates, a substantive theoretical advance with moderate field impact.
In briefStudy on Directions of Reasoning Correctness in Models
SummaryThe paper introduces ChainSearch, a controlled arithmetic reasoning task extending Countdown with explicit search and backtracking, and trains a 3-million-parameter model on it. Using linear probes, mean activation differences, and direct optimisation, the authors search for internal reasoning-correctness directions. They find statistically significant such directions, but none consistently improves reasoning under causal intervention. The results question whether the linear representation hypothesis suffices to identify causally useful directions for reasoning.
Munachiso Samuel Nwadike, Zayd M. K. Zuhri, Rifo Genadi et al.
In briefDiaLong: A Benchmark for Memory in Long-Context Dialogues
SummaryThis paper introduces DiaLong, a long-context dialogue dataset and memory task for measuring how well language models retain and recall information in long conversations. The dataset contains 100 human-created conversations augmented with LLMs, and the task asks models to classify 40 candidate facts about each dialogue as true or false. Testing the latest LLMs and memory enhancement methods shows that current models still need better memory performance.
In briefCompressing medical segmentation models while keeping uncertainty reliable
SummaryThis paper studies whether compressing pretrained medical image segmentation models preserves uncertainty signals, not just Dice scores and model size. The authors propose SGAV-MP, a post-training method that chooses per-layer quantization precision using a benefit-risk score and accepts changes only after full-model reliability validation on a calibration set. Reported results across five 2D backbones and three 3D benchmarks show FP32-level Dice and high uncertainty-map agreement, with compression of about 3.59 to 7.8 times in 2D and 4.73 to 7.82 times in 3D.
Sagar Kumar, Shailja Sharma, Tirthajit Baruah et al.
In briefBiDex-100: A benchmark testing two-handed dexterous robot manipulation across over 100 simulated tasks
SummaryBiDex-100 is a simulation benchmark for robots that use two dexterous, multi-fingered hands. It contains 106 tasks grouped under five capabilities: generalization, instruction following, limited demonstrations, long horizon, and memory, with out-of-domain variants that change visuals and spatial layout. The benchmark supports two robot embodiments and provides 8,080 success-validated teleoperated demonstrations totaling 6.25 million frames. The authors evaluate seven recent Vision-Language-Action models under a unified protocol to establish baselines.
In briefTraining Small Phone Models to Answer and Summarize Calls
SummaryThis paper builds training data for small language models that screen calls on a phone for an absent owner, keeping call content on the device. Using synthetic calls checked by model-based reviews, it assembles 3,395 calls with 15,102 training targets. After training a shared adapter, reply and summary qualification rates both rise on 192 held-out calls, but these rates reflect agreement with one evaluator's standard, not human judgment.
In briefSystem X: An LLM decision model that answers from its own KV cache
SummaryThis paper presents System X, an LLM decision model that adds a text-reply option, continuing generation from the decision's KV cache to avoid re-prefilling the prompt. It reports no significant change in JevBench accuracy, lower calibration error, and 15-37% latency savings on an Apple M4 Pro. The v4 version using Qwen3-1.7B improves JevBench accuracy, while state conflicts and unsupported reply details remain failure modes.
Syed Usama Khalid Bukhari, Syed Safwan Khalid, Muhammad Aamir
In briefSemi-supervised learning on singular data: modeling irregular data clouds with subanalytic sets
SummaryThis paper models singular data point clouds using subanalytic sets and develops a semi-supervised learning approach in this setting. It relies on a Morrey-type embedding for Sobolev spaces with respect to globally subanalytic measures that may have singular support. The authors present a greedy method that uses a Monte Carlo approximation of the Sobolev norm of candidate solutions. Numerical experiments are reported on point clouds sampled from real algebraic varieties and on a classification task using MNIST images downsampled to 7×7.
In briefUBD-Bench: Benchmarking Weakened Single-Sensor Branches in Multimodal Self-Driving Perception
SummaryThis paper introduces UBD-Bench, a benchmark for studying how the single-sensor branches of multimodal self-driving perception models can degrade. It uses a Paired Frozen-Branch Probe to compare the same modality's representations after unimodal and multimodal joint training. Across four perception tasks and 10 models, camera branches consistently degrade after joint training, while LiDAR branches change less and depend more on the model. Distilling from unimodal to multimodal training at the degraded representation improves camera-branch probe performance and full-model performance on all four tasks.
In briefEnhanced Masked Diffusion Unlearning (EMDU): Making Masked Diffusion Language Models Forget Target Content
SummaryMasked diffusion language models generate text by iteratively unmasking tokens using bidirectional context. The paper identifies two failure modes of the prior method MDU and proposes EMDU, which uses a frozen anchor with the entire answer span masked and an anchor-first KL divergence. On the TOFU and RWKU benchmarks using LLaDA-8B and Dream-7B, EMDU matches or exceeds MDU and adapted autoregressive baselines in forgetting while better preserving retained utility.
In briefStepwise Reasoning Verification: Classify Steps, Don't Generate Verdicts
SummaryThis paper treats stepwise verification of large language model reasoning as a natural language inference task, classifying whether each next step is entailed by the problem and partial solution. Across five reasoning datasets, a 0.15B encoder classifier outperforms most generative verifiers with up to 4B parameters. On identical decoder backbones, classification beats generation from 0.5B to 4B and matches it at 7B, using roughly one tenth as many annotated problems. The authors conclude that strong verification does not require large generative decoders.
In briefDesign Choices for OCR on a Low-Resource Indigenous Language Archive
SummaryThis paper examines how small vision-language models with up to 3B parameters can be adapted as OCR tools for Lakota, a North American Indigenous language. It considers practical issues such as limited training budgets, segmentation noise, scan degradation, and spelling variation. Fine-tuning on transcribed pages for OCR post-correction substantially reduces character error rate on a bilingual corpus, and training with fewer pages costs little accuracy. However, under realistic automatic segmentation, the best models' error rate nearly doubles, highlighting the need to reflect real use cases.
T\'ea Wright, Laura X. Ma, Claudia Iron Hawk et al.
In briefCan PRNU still identify modern smartphones? Evidence that device fingerprints survive computational photography
SummaryThis paper examines whether PRNU camera fingerprints can still identify modern smartphones, which recent studies found to show high cross-device correlations. Using 30 recent smartphones under controlled conditions, the authors find that a component shared by same-model fingerprints inflates the standard sPCE score. They propose ΔsPCE, which subtracts the mean sPCE against the other candidate fingerprints, improving attribution accuracy. The authors note limits from the small dataset and dependence on the candidate camera set, but conclude that PRNU-based attribution remains feasible when fingerprints are compared appropriately.
In briefDragonMoCap: Video-Based Motion Capture for Many Dragon Rig Layouts
SummaryThis paper addresses motion capture for dragon characters from single-camera video, a task that currently depends on manual keyframing because suitable datasets and parametric models are lacking. The proposed DragonMoCap framework treats the target rig as a dynamic input and uses language-grounded joint query tokens with a non-parametric retrieval mechanism, allowing zero-shot adaptation to unseen rigs without retraining. The authors report that it outperforms HMR-style baselines across diverse dragon topologies and has been integrated into professional animation workflows.
Zhongfei Qing, Xiaoxiang Zhao, Ling-Hao Chen et al.
In briefRARE: Reinforcement Learning with Adaptive Rubric Evolution for Open-Ended Generation
SummaryThis paper proposes RARE, a reinforcement learning method for open-ended generation tasks that lack automatic verifiers. Based on current policy feedback, it replaces criteria that are fully satisfied and no longer discriminative with checks of remaining quality differences, keeping other criteria unchanged. On video script generation across three video types, it achieves a mean gain of 3.82 points over matched static rubrics, supported by model and human preferences. The paper also reports improvements on writing and general generation benchmarks.
In briefMOSAIC: Searching for Per-Layer Memory Budgets in Large Language Models
SummaryMOSAIC is a multi-objective neural architecture search that learns how much KV-cache memory each transformer layer should get, for use with existing token eviction methods. On Llama-3-8B-Instruct and Mistral-7B-Instruct-v0.2, it beats uniform allocation in 146 of 154 settings, with gains up to 48 points on RULER. It reaches uniform-allocation scores with up to 4x less KV cache, and the learned allocations consistently favour a small set of middle layers.
In briefPreserving SAR Structure: Frozen Expert Alignment for Label-Efficient Radar Scene Understanding
SummaryThis paper aligns synthetic aperture radar (SAR) images with natural language by keeping both a SAR encoder (CROMA) and a CLIP text encoder frozen and learning only a lightweight constrained residual bridge. The authors measure physical forgetting with Procrustes disparity, reporting that their bridge preserves SAR geometry better than fine-tuning (0.210 vs. 0.351). On zero-shot SAR scene classification, the method outperforms fine-tuned baselines while using up to 8 times less training data.
Eli A. Underwood, I. Taylor McKechnie, Leslie M. Collins
In briefMixture-of-Memories: a low-overhead way to add modular knowledge to language models
SummaryThis paper addresses the difficulty of controlling, updating, and reliably retrieving facts stored in large language models. It proposes Mixture-of-Memories (MoM), which separates routed neural memory from writable factual memory, using a frozen base model with latent and lexical memory experts and a lightweight router that sets per-token weights. Experiments on language modeling, factual QA, update analysis, and efficiency indicate improved predictive accuracy and updateability with low inference overhead.
In briefFrom Label Priors to Task Evidence: Efficient Long-Video Frame Selection via Bayesian GRPO
SummaryThis paper addresses how multimodal language models can efficiently process long videos by selecting key frames. It first trains a lightweight frame selector with supervised learning, then refines it using a frozen multimodal model's answer likelihood as a reward within Bayesian GRPO. Experiments on Video-MME, MLVU, and LongVideoBench report competitive accuracy with lightweight efficiency, and the selector transfers to multiple downstream models without retraining.
WhyReframes frame selection as latent-variable inference, using frozen MLLM answer likelihood as dense reward, validated on three long-video benchmarks with cross-model transfer.
In briefUniState: Unified memory and intervention for controllable world simulation
SummaryThis paper introduces UniState, a framework for controllable world simulation that uses one unified world-state representation combining spatial memory and user interventions. Users specify object motion within spatial memory and separately choose the camera trajectory used to observe the resulting state. Training and sampling strategies help generated videos follow the specified object motion and camera path while still producing plausible additional responses. Experiments report more accurate joint object and camera control than existing methods.
Jaeseok Jeong, Kyungmook Choi, Sohyun Chung et al.
In briefEvaluating how well large language models understand figurative language in Moroccan Darija
SummaryThis paper introduces the Moroccan Darija Figurative Language Corpus (MDFC), a benchmark of idioms, metaphors, and proverbs for machine translation and multiple-choice tasks. The authors evaluate 19 open- and closed-weight language models and translation systems at expression and contextual levels. Closed-weight models achieve the strongest translation performance, while the value of context varies across models and tasks.
Jihad R'baiti, Youssef Hmamouche, Emmanuele Chersoni et al.
In briefMT Forcing: Stable, Consistent Multi-Turn Image Editing via Autoregressive Diffusion
SummaryThis paper addresses drift in multi-turn image editing, where visual quality degrades and unintended content changes appear after a few instructions. It proposes MT Forcing, a training paradigm for unified multimodal models based on autoregressive diffusion. The method combines Single-Turn Distribution Matching Distillation, which guides student rollouts with a teacher applying cumulative edits in one turn, and Teacher-Forcing Guidance, which supervises cross-turn transitions. The authors report that experiments show more stable visual quality and stronger cross-turn consistency.
In briefWhen Form Changes but Logic Doesn't: Logic-invariant LLMs via Logical Structure
SummaryThis paper examines whether large language models truly understand logical structure, finding that they often give inconsistent predictions for logically equivalent inputs and generalize poorly to unseen logical forms. The authors propose LoGIST, which maps each input to a decision diagram, a directed acyclic graph capturing its logical structure, and encodes it as an embedding that guides the model's reasoning. Across multiple training settings, LoGIST reduces inconsistency under logical equivalence and improves generalization to unseen logical forms.
In briefRAHF-Net: A range-aware network that fuses LiDAR and camera images for better pixel-level scene labeling
SummaryThis paper studies how LiDAR point-cloud features can improve pixel-level semantic segmentation of RGB images. The authors propose RAHF-Net, which projects pretrained cylindrical voxel features into the image plane and uses a depth-derived soft range mask to apply LiDAR-guided enhancement in near-field regions and visual-context enhancement in sparse far-field regions. On KITTI-360 it reaches 65.76 mIoU, 1.45 points above CMX, and experiments on DELIVER report improved robustness under adverse weather, illumination degradation, and sensor corruption.
In briefTurning Unverifiable English Domain Text into Cross-Lingual RL Training Tasks via Code-Switching
SummaryThis paper proposes a data-centric method that turns English domain text, which cannot be directly verified, into verifiable reinforcement learning tasks in target languages and in code-switched (mixed-language) forms. It uses fill-in-the-middle multiple-choice questions as the seed task format. Experiments on Korean medical benchmarks and cybersecurity benchmarks across four target languages show that code-switched training outperforms monolingual and bilingual-mixture baselines on average target-language performance, while English performance remains comparable.
In briefPolarVis: RGB-to-Polarization Perception Dataset and Diffusion Framework for Driving Scenes
SummaryThe paper introduces RGB-P Wild, a dataset of 10,000 paired RGB and four-angle linear-polarization image sets, and PolarVis, a diffusion-based framework that recovers polarization representations from a single RGB image. On this dataset, PolarVis reaches a PSNR of 27.83 dB and an SSIM of 0.8597. In downstream detection, a frozen PCDNet using PolarVis-generated polarization scores 86.3%/59.1% (AP50/AP50:95), close to 87.0%/59.8% with measured polarization.
In briefA 3D-driven framework that turns moving-camera video into 360° video
SummaryThis paper presents a modular framework that converts a perspective video shot with a moving camera into a 360-degree video. Input frames are reconstructed explicitly in 3D, while a pretrained panoramic image model fills in unobserved regions, and both are combined in a shared 3D scene to stay consistent across viewpoints. The authors report state-of-the-art input-view fidelity, perceptual quality, and geometric consistency on a static-scene subset of the Argus evaluation set, without additional panoramic-video training.
Michael Wakeham, Mrinmoy Kumar Das, Wenqi Wang et al.
In briefPaloma: Phase-conditioned residual modulation for more accurate time series forecasting
SummaryThe paper argues that residuals left after removing periodic components in time series still contain structured, phase-dependent deviations rather than pure noise. It proposes Paloma, which combines a Phase Rotation module and an Amplitude Scaling module to model these directional and magnitude changes. The Paloma technique is packaged as a plugin that can be added to existing forecasting models, and experiments on twelve datasets report state-of-the-art and robust results.
WhyTreating residuals as phase-conditioned signals with rotation and scaling is a substantive improvement; validated on twelve datasets as a general plugin.
In briefATLAS: Attention-guided search for better compositional image generation
SummaryThis paper introduces ATLAS, a training-free method that uses a frozen image generator's internal attention to guide the search among candidate generation paths before final images are produced. It keeps several coarse alternatives, prunes candidates at intermediate scales without decoding, and evaluates images only for a final choice among partial trajectories. The authors report higher scores on T2I-CompBench and GenEval and over 50% less generation time than completing all candidates.
Hossein Shahabadi, Raya Namazi, Mahdieh Soleymani Baghshah
In briefHidden Look-Ahead: Timing Errors in Electricity Market Forecast Data
SummaryThe paper examines whether electricity price models that use commercial weather forecast archives truly rely only on information available at trading time. Checking timestamps of 128,656 forecast runs from ten weather-model configurations against Great Britain's day-ahead and intraday auction gates, the authors find that the "latest run issued before the gate" was often not yet published at the gate. They propose a protocol that audits point-in-time correctness from timestamps before any model is fitted.
In briefOnly Means Only: what a harmonization benchmark score averages over
SummaryThis paper finds that about a quarter of the test composites in an image harmonization benchmark come from photographs also used in training, despite the benchmark's no-overlap promise. It shows that the reported scores average over two populations mixed in an unreported proportion. The authors release a list of 5,581 test composites without training-set construction identities, usable without retraining.
In briefStructured relevance and channel-additive Gaussian processes for images
SummaryThis paper proposes radial basis function kernels for Gaussian processes on image inputs, using structured sharing of lengthscales across pixel patches and colour channels. This reduces the number of independently learned hyperparameters compared with fully anisotropic kernels. It also introduces channel-additive Gaussian processes for multi-channel images. The kernels are tested on MNIST and a small RGB satellite aircraft-detection dataset, where the authors report better predictive performance than isotropic and fully anisotropic baselines.
In briefOutcome-Anchored Credit Reallocation: a training method that better assigns rewards for multi-step knowledge graph question answering
SummaryThis paper addresses agentic knowledge graph question answering, where language models interact with a knowledge graph over multiple turns. It notes that standard outcome-supervised reinforcement learning gives every generated token the same outcome advantage, which mismatches the varying effects of individual knowledge graph operations. The proposed method, OACR, uses a training-only knowledge graph state outcome model to estimate reward changes between search states and converts them into token-wise weights that adjust update magnitudes. Across three benchmarks, OACR reports stronger overall performance than existing methods while using fewer inference-time LLM calls and tokens.
In briefWhen weak submodularity meets matroids: how well does standard greedy perform?
SummaryThis paper studies the worst-case approximation guarantee of the standard forward greedy algorithm for weakly submodular maximization under general matroid constraints, where γ is the submodularity ratio and r is the matroid rank. The authors derive a refined lower bound that strictly improves the previous 0.4γ²/(1+√(γr)) guarantee and recovers the classical 1/2 bound at γ=1. They also give a matching-style upper bound that sharpens the known O(1/log(n-1)) bound to O(r^{-1/2}). Together these yield a tight characterization of the approximation ratio up to constants.
WhyResolves a longstanding open question on greedy under matroid constraints with near-tight, rigorously proven bounds improving prior guarantees.
In briefSimulating infant senses with age-tuned neural networks
SummaryThe authors combine existing measurements of infant vision, hearing and touch over the first two years into age-indexed input filters, with a small prediction network learning behind each. The filters alone reproduce the order in which the senses mature. The networks show that early vision extracts far less than its input allows (R² of 0.00 at birth for a toy's position, versus 0.57 for a network trained directly on the same input), while hearing tracks what its input provides. The companion Baby Senses Lab website shows users their own images and sounds as a baby of a chosen age might sense them. It is intended as a research and teaching tool, not a diagnostic test.
In briefTILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models
SummaryThis paper presents TILDE, a method that transfers the manipulation skills of large pretrained Embodied Foundation Models to policies that did not undergo action pretraining. It combines on-policy distillation on states induced by the student with learning from teacher-executed recovery trajectories. The authors also introduce AutoDAgger, which uses a robotic reward model to monitor task progress and trigger teacher intervention, collecting recovery data without human monitoring and more than 2 times faster than a human-in-the-loop approach. Across three real-world tasks, TILDE improves task performance by up to 4.1 times over behavior cloning and supports cross-architecture distillation.
In briefMining Self-Rewriting AI Systems: A Data-Mining Agenda for AI-Native Infrastructure
SummaryThis paper introduces the "self-rewriting system," in which agents' own actions reshape the logs, corpora, and records later used for training and evaluation, causing what the authors call endogenous drift. It argues that monitoring each data source separately cannot reliably attribute causes across sources, so joint analysis is required. The authors outline four research problems and propose public benchmarks, using AI-native telecom as the motivating example.
In briefMemDerive: A Diagnostic Benchmark for Agent Memory over Evolving Data
SummaryThis paper introduces MemDerive, a diagnostic benchmark for testing how well agent memory handles evolving data. It defines scenarios symbolically, computes exact gold answers with Datalog, and records the derivation graph and supporting events for each query. The benchmark includes 10 scenarios, 4,692 events, and 1,503 queries. The evaluation finds that memory systems fall behind simple BM25 retrieval on derived state, mostly because they fail to retrieve needed intermediate states.
In briefWhistles and wrong lyrics: synthetic data alone does not solve real query-by-melody
SummaryThis paper studies searching for songs by whistling or singing with incorrect lyrics, rather than humming. The authors train on synthetic whistles and nonsense-syllable singing from three generators, but these gains do not transfer to real whistles. Fine-tuning on real whistle recordings improves top-1 retrieval, so the authors conclude that collecting real queries is the missing step.
In briefMulti-site deep learning for day-ahead wind power forecasting in Great Britain
SummaryThis paper trains multi-site deep neural networks on historical production data and reanalysis weather from 88 onshore wind units in Great Britain to forecast next-day wind power output, matching NESO's day-ahead schedule. A recurrent network with a bidirectional weather encoder (BiRNN) performed best, lowering mean absolute error and root mean squared error relative to NESO by 17.94% and 12.51% across 638,246 matched hourly observations. In leave-one-out trials, it beat NESO's error at 32 of 40 excluded sites.
In briefPainting Change Map: guiding video generation of painting processes
SummaryThis paper introduces the Painting Change Map, a spatio-temporal control mechanism for guiding video generation of both forward painting and reverse unpainting. It tracks normalized structural changes over 1,000 timesteps and encodes localized visual differences into RGB space, specifying where and when an artist acts. The authors report that their method outperforms existing control baselines and achieves competitive or superior results compared with state-of-the-art painting methods.
Markus Pobitzer, Chang Liu, Niccol\`o Biondi et al.
In briefAligning Language Models with Arab Cultural Behavior Using Preference Data
SummaryThis paper addresses the gap in aligning Arabic large language models with country-specific cultural behavior, which prior work largely overlooked in favor of factual instruction tuning. The authors introduce ArabPref, which they describe as the first large-scale pan-Arab preference dataset for cultural behavior conditioned on country. They release a preference dataset for training and evaluation, and a multiple-choice benchmark, in English and Modern Standard Arabic, with test data validated by native speakers. They experiment with post-training methods on multilingual and Arabic-centric models.
In briefOperator-Mismatch Contrastive Learning for World-Relative Trajectory Coherence
SummaryThis paper argues that world models should be judged not only by how plausible their predicted futures are, but by whether they follow the latent rules actually in force. The authors propose Operator-Mismatch Contrastive (OMC) learning, which uses valid-but-wrong-operator trajectories as hard negatives, along with a protocol that measures world-relative coherence directly. Experiments on a synthetic benchmark and feature-level external validation with CoPhy ballsCF show that OMC substantially improves operator-sensitive scoring.
One Octadion, Firmansyah Sundana, Novanto Yudistira
In briefWho Persuaded Whom? How LLMs Change Their Minds in Multi-Agent Debate
SummaryThis paper studies how peer arguments shape individual and group decisions when multiple language models debate each other. The authors introduce MAD Bench, a 2,400-question benchmark covering knowledge and reasoning, values and moral judgment, and safety and refusal, and compare debate against an independent-reflection control. They find that agreement after debate does not necessarily reflect independent corroboration, and that debate can increase unanimous wrong answers, motivating evaluations that also assess dependence and shared errors.
Arnesh Batra, Aniket Khandelwal, Jashn Khemani et al.
In briefCD-OPD: Contextual Digest Method for On-Policy Distillation
SummaryThe paper studies on-policy distillation (OPD), where a teacher model can take over and cause wrong answers even when both teacher and student solve the same problem correctly, measured by Collaboration Failure Rate (CFR). The proposed CD-OPD method uses a teacher-generated dependency graph to summarize problem conditions and prior reasoning into a Contextual Digest that the student continues from. It also adds online meta-reweighting to guide updates toward reducing task loss. Across eight benchmarks, the authors report accuracy gains over representative OPD methods and lower CFR than takeover-based approaches.
In briefLateMatch: Cacheable Matching to Speed Up Visual Localization on Edge Devices
SummaryThis paper addresses the high computational cost of deep feature matchers used for visual localization on edge devices. It proposes LateMatch, which defers inter-image interaction to the final layers so intra-image processing can be cached offline. The authors also introduce E-SToRE, which uses decreasing Shannon entropy to reduce token counts before cross-image interaction. The paper reports over 2x lower effective computational cost than standard matchers, with E-SToRE saving an additional 10% to 14% and negligible accuracy loss.
In briefCHSM8: A Chess Model That Plays in Styles Described by Text
SummaryThe paper presents CHSM8, a 210M-parameter chess transformer that chooses moves based on a free-text description of a playing style, such as a named opening or a player's manner. It is pretrained without search on 1.07B filtered Lichess games. After fine-tuning on a synthetic text-labeled dataset and direct preference optimization, style adherence improves across tested prompts while playing strength without a prompt stays close to the supervised baseline.
In briefLocal Visual Evidence Versus Full-Video Failure in Audio-Visual Models
SummaryThis paper studies why audio-visual large language models can fail at full-video speech-vision judgments even when useful visual evidence is locally recoverable. It introduces TED, a training-free diagnostic framework that separates local visual verification from aggregation relative to the speech interval. Experiments on SVHalluc with four frozen models report higher temporal forecasting accuracy with TED than with full-video questioning, alongside control tests for alternative explanations.
In briefWhen Planning Problems Become Hard for LLMs: Separating Structure from Semantics
SummaryThis paper examines how large language models perform on classical STRIPS planning problems, separating a problem's structural difficulty from the difficulty of interpreting symbolic vocabulary. It finds that the zero-training delete-relaxation heuristic h_add predicts structural difficulty better than the learned alternatives tested on held-out domains. When predicates are progressively renamed while states, transitions, and optimal plans are preserved, LLM planning success falls markedly, while refusals stay at zero. The authors conclude that improving hybrid planning requires modeling LLM sensitivity to predicate semantics rather than building more structural predictors.
In briefPredicted Class Correction: Conformal Thresholds for Classes Lacking Calibration Data
SummaryThe paper proposes Predicted Class Correction (PCC), a method that sets per-class conformal prediction thresholds for classes with no labelled calibration examples. It predicts the offset from the marginal threshold using geometric descriptors of the classifier head, which need no labels, and calibrates only a single scalar. On ImageNet, PCC improves worst-class coverage for such classes at matched average set size. The authors also report that worst-class coverage cannot be reliably measured with fewer than roughly 35 test examples per class.
WhyPredicts conformal thresholds for unlabelled classes from classifier weight descriptors, with baseline comparisons and an evaluation-depth analysis providing solid evidence.
In briefMCSplat: Consistent Feed-Forward Gaussian Splatting for Multi-Camera Driving Scene Reconstruction
SummaryThis paper presents MCSplat, a feed-forward Gaussian splatting framework that reconstructs driving scenes from sparse multi-view images. It builds a canonical appearance space from depth-guided multi-camera correspondences to separate geometry learning from camera-specific exposure and illumination differences. A multi-scale Bilateral Grid Transformer restores raw camera appearance for photometric supervision, and Fisher-information uncertainty weights surface self-supervision. Experiments report improved reconstruction quality, generalization across datasets, and better cross-view depth and surface-normal consistency.
In briefFoundPDE: A pretrained model that discovers partial differential equations from data
SummaryFoundPDE is a pretrained transformer model that takes discretized spatio-temporal trajectories and outputs symbolic partial differential equations. On a clean benchmark covering 20 PDE families, it reaches 97.50% type recognition, which drops to 52.09% under 10% multiplicative Gaussian noise. For unseen composite equations, few-shot fine-tuning achieves over 90% exact recovery on four held-out families.
In briefCASE: A complementary speech encoder that better matches lip movements for speech-driven 3D face animation
SummaryThis paper introduces CASE, a speech encoder adapted by contrastively aligning short speech clips with continuous 3D lip-mesh latents from a VQ-VAE, so its representations better reflect visible articulation. Fused with a baseline self-supervised speech representation across four VOCASET baselines, CASE lowers mean lip vertex error, and mean full-face vertex error decreases in every system. Adaptation used Korean speech-geometry pairs while evaluation used English VOCASET, consistent with cross-language downstream transfer.
In briefFormally verified AI-driven research on how the number of attention heads limits what transformers can compute
SummaryThis paper studies how many attention heads a one-layer, attention-only transformer needs to represent a Boolean function, defined as its head complexity. The authors derive an exact normal form, a general threshold-degree lower bound, and an explicit ten-bit function where that bound is strict. For symmetric Boolean functions, they show head complexity equals the number of value changes between consecutive Hamming-weight levels. A frontier model generated and machine-verified the theorems, with humans posing questions and reviewing results, and the Lean formalization is publicly released.
WhyGives an exact characterization of attention head complexity with a threshold-degree lower bound, fully machine-verified, though impact is mainly theoretical and interpretability-focused.
Karthik Viswanathan, Plamen Dimitrov, Stefano Rocca et al.
In briefLightweight Real-Time Driver Safety Framework for Edge Devices with Lane Departure and Traffic Sign Recognition
SummaryThis paper presents a unified driver-assistance framework that runs on automotive-class edge hardware and performs both lane-departure warning and traffic-sign detection and recognition. The lane warning is based on lightweight lane segmentation with a drivable-area fallback, and the sign branch uses a two-stage pipeline trained on a curated 35-class local dataset expanded to 8,750 samples. On an NVIDIA Jetson Orin Nano, the system runs at 15 FPS end to end, with 97.4 mAP for detection and 95.6% validation accuracy.
Hassaan Ali Qureshi, Danish Hussain, Ali Hassan et al.
In briefA distribution-free separation showing learned features can beat fixed ones
SummaryThe paper shows that learned features can offer a computational advantage over fixed features even without assuming any input distribution. The authors build a hypothesis class that a two-layer ReLU network trained with a natural gradient-based algorithm can learn efficiently, but that linear predictors over any polynomial-dimension fixed feature map cannot represent. This settles an open question posed at COLT 2026 by Feldman, Kamath, and Srebro.
WhySettles an open question by constructing a distribution-free theoretical separation between learned and fixed feature representations.
In briefCoRe-WM: A world model for predicting breast cancer treatment response from longitudinal MRI
SummaryThis paper addresses predicting pathologic complete response (pCR) after neoadjuvant chemotherapy for breast cancer by framing it as a world-modeling problem. The proposed CoRe-WM learns a latent model that predicts the next imaging state from an observed sequence, conditioned on patient clinical profile and assigned therapy, with a factorized pathway separating response dynamics from imaging appearance. Evaluated on I-SPY2 with five-fold cross-validation and transferred to I-SPY1, it achieves the highest discrimination at every prospective decision point among compared methods.
In briefFollow the Mixed Groups: Sampling Method That Trains Web Agents on Informative Questions
SummaryThe paper observes that group-relative policy optimization learns only when sampled answers disagree, since all-correct or all-incorrect groups give zero advantage. It proposes RIDGE, which keeps a Gaussian curriculum but weights prompts by Beta posteriors and moves the curriculum center and width only from groups that come back mixed. On two math backbones and a retrieval-augmented question answerer, RIDGE raises average accuracy over GRPO by 3.4 and 2.9 points and average exact match by 3.7 points. It also uses fewer math rollouts and tool calls than dynamic sampling.
In briefMPIG: Generating Multi-Angle Polarization Images from One Polarization View
SummaryThis paper proposes MPIG, a Swin Transformer-based framework that predicts the missing 45°, 90°, and 135° polarization images from a single 0° observation. The authors also introduce HighPol, a real-world dataset of 5,474 groups of high-resolution four-angle polarization images from indoor and outdoor scenes. The reported experiments show improved image and polarization accuracy, and downstream glass and semantic-segmentation tests suggest the generated images keep task-relevant information.
In briefPaper overview: a probe-based prior for finding prompts whose sampled groups split in RL post-training
SummaryThis paper addresses a waste in group relative policy optimization: when all sampled rewards in a group agree, within-group variance and advantage are zero, so the rollouts teach nothing. The authors propose Aperture, which trains a small subset with the same recipe, scores prompts on the base and probe models, and uses the two pass rates to estimate the chance a training-size group is all wrong or splits, as a static sampling weight. On mathematical and logical reasoning with 1.7B and 4B models, the prior improves greedy accuracy over uniform GRPO and increases the fraction of groups with nonzero advantage.
In briefWhy search-augmented LLMs fail when reasoning over fast-changing knowledge
SummaryThe paper introduces TREK, a suite with TREKTRACE, a framework that attributes failures to specific retrieval or reasoning stages, plus two benchmarks: TREKBENCH-WILD for real-world questions and TREKBENCH-CTRL with controlled distractors. The authors find the main bottleneck is post-retrieval reasoning, as frontier models are often misled by conflicting evidence even after retrieving the needed information. They also show that benchmarks dominated by outdated distractors substantially underestimate task difficulty.
