Daily Summary for 2026-11-12
daily
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: It's the twelfth of November, twenty twenty-six, and this is the day's research.
Jane: 49 new papers came out today.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We'll take the day in one pass, then pull out the papers we're staying with.
Jane: Start us off with LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution.
The day's research: Tom: Today's focus is on extracting structured information from visual data which is crucial because scientific figures are often the primary way research claims are communicated.
Jane: This is addressed by SciFigure2Code, which builds an AI-reconstructed benchmark to test presentation recovery by generating editable Python programs.
Lu: This reconstruction protocol involves role-specialized Codex agents that generate, refine, and audit these programs creating an auditable reference package for published panels.
Meng: While execution fidelity remains weak across models when testing the SciFigureBench set, Claude Opus 4.7 shows the best performance in image-only reconstruction.
Lalam: Using plan-then-code prompting helps all tested models improve, which is a key finding for that benchmark set.
Tom: Moving to representation learning, RepFusion explores how large language model priors can be leveraged for denoising visual representations by repurposing multimodal LLMs as a noisy representation encoder.
Jane: This method suggests that MLLMs provide strong priors for denoising visual representations, outperforming baselines that use newly initialized denoisers at similar inference budgets.
Lu: The work on video generation touches on the need for foresight in autoregressive models, as Video-Mirai introduces a training-only method where a frozen foresight encoder reads the completed rollout to distill causal states into causal states.
Meng: This shows that visual autoregressive models require this kind of future supervision to maintain consistency over longer rollouts.
Lalam: Finally, we look at improving image quality in degraded scenarios with LucidNFT, which uses a multi-reward reinforcement learning framework for flow-matching real-world super-resolution.
Tom: This method introduces LucidConsistency to detect hallucinations by training an evaluator on content-consistent degradation pools and has shown its outputs are judged more faithful to the original LR inputs in seventy two point five percent of human studies.
Jane: The most significant development is Causal Attribution via Activation Patching because it offers a principled way to determine which specific parts of a vision transformer are responsible for its decisions.
Lu: This matters because existing methods often provide only indirect proxies for patch importance, whereas this technique captures the causal contribution of internal representations after they have formed.
Meng: This approach is supported by work like Spatially-Aware Class-Agnostic Object Counting, which improves structural sensitivity in object counting by reassembling multi-layer features into a refined pyramid.
Lalam: This refinement builds upon the need for better spatial modeling that is also evident in CD-RCM, which tackles the anisotropic nature of reflectance confocal microscopy stacks by introducing a feedforward model.
Tom: Furthermore, the integration of cognitive elements is being explored in Mixture of Cognitive Experts in Large Vision-Language Models, where a Bloom-inspired taxonomy is used to convert expert outputs into a staged reasoning trace.
Jane: This structured reasoning is relevant when dealing with complex visual and textual questions that require diverse cognitive levels.
Lu: Finally, for autonomous systems, Generative Scenario Rollouts use autoregressive rollout strategies to generate language-grounded future traffic scenes, which helps stabilize predictions through a rollout-consistency loss.
Meng: This generative planning capability is also relevant to agents in dynamic settings as seen in 3A-VLA, which grounds action learning by using intention and environment abstractions to guide grounded actions.
Lalam: SciFigure2Code: An AI-Reconstructed Benchmark for Scientific Figure-to-Code Scientific figures are hard to convert into code, so this tool reconstructs them to test AI's ability to reproduce scientific plots.
Tom: RepFusion: Leveraging Multimodal Priors for Denoising in Representation Space This method uses large language model outputs as a signal to help denoise visual representations during image generation.
Jane: LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution This technique uses reinforcement learning to improve the quality of super-resolution images from low-resolution inputs by focusing on faithfulness.
Lu: Video-Mirai: Autoregressive Video Diffusion Models Need Foresight Causal video generators need a way to look into the future to maintain consistent identity and motion in their generated sequences.
Meng: A squared: Smaller Self-Supervised ViTs Localize Better than Larger Ones This method combines small models for localization with large models for rich features to get the best of both worlds.
Lalam: Are Text-to-Image Models Inductivist Turkeys? A Counterfactual Benchmark for Causal Reasoning This benchmark tests if text-to-image models truly understand cause and effect by asking them to generate images under contradictory rules.
Tom: Lighting-aware Unified Model for Instance Segmentation Foundation models struggle with varying lighting, so this model uses contrast maps to make segmentation robust to different illumination.
Jane: DVFace: Spatio-Temporal Dual-Prior Diffusion for Video Face Restoration This method uses dual spatial and temporal priors in diffusion models to restore high-quality, temporally consistent faces from degraded videos in a single step.
Lu: Causal Attribution via Activation Patching This technique identifies which specific image patches cause a Vision Transformer model's prediction by directly looking at its internal activations.
Meng: Mixture of Cognitive Experts in Large Vision-Language Models This framework uses a hierarchical reasoning protocol to integrate different cognitive experts into one coherent representation for better decision-making.
Lalam: A Generative Data Framework with Reliable Supervision for Underwater Image Restoration and Enhancement This approach creates reliable training data for underwater image restoration by synthesizing missing parts using real source images as references.
Tom: 3D-Guided Scalable Flow Matching for Generating Volumetric Tissue Spatial Transcriptomics from Serial Histology This method uses 3D flow matching to accurately predict gene expression in unprofiled tissue sections based on neighboring sections.
Jane: Generative Scenario Rollouts for End-to-End Autonomous Driving This framework uses autoregressive generation to plan and generate consistent future traffic scenes for autonomous driving models.
Lu: Spatially-Aware Class-Agnostic Object Counting This method uses multi-layer features and a specialized head to count objects in images while focusing on preserving their spatial structure.
Meng: CD-RCM: Generalizable Continuous-Depth Novel View Synthesis for Reflectance Confocal Microscopy This model synthesizes realistic, unseen depths from sparse 3D stacks of microscopy images to create continuous volume representations.
Lalam: 3A-VLA: Abstraction-Aligned Action Learning for Vision-Language Agents in 3D Game Worlds This framework uses explicit abstractions to help vision-language agents learn better actions in complex 3D game environments.
Tom: Deep Learning-based Filtering for Video Coding: A Survey on Architectures, Algorithms, and Complexity Analysis This survey categorizes deep learning filtering methods for video coding based on hardware feasibility and performance trade-offs.
Jane: BigEarthNet.txt: A Large-Scale Multi-Sensor Image-Text Dataset and Benchmark for Earth Observation This dataset provides a large collection of multi-sensor satellite images with rich text annotations for remote sensing tasks.
Lu: VL-UniTrack: A Unified Framework with Visual-Language Prompts for UAV-Ground Visual Tracking This framework uses visual language prompts to create a unified framework that tracks targets between UAV and ground views.
Meng: CRIL-U-Net: Compact Ratio-Interaction Learning for Focal Cortical Dysplasia Segmentation from T1w and FLAIR MRI This model uses a compact learning module to improve the segmentation accuracy of small focal cortical dysplasia lesions in MRI scans.
Lalam: PrefReward: Learning User Preference Matrix for Personalized Text Generation This framework learns user styles by creating a preference matrix and using it as a reward signal to guide personalized text generation.
Tom: Matching-Based Few-Shot Semantic Segmentation Models Are Interpretable by Design This method makes few-shot segmentation models interpretable by design through affinity scores derived from their support sets.
Jane: CLUE: Controllable Latent space of Unprompted Embeddings for Diversity Management in Text-to-Image Synthesis This framework allows text-to-image models to generate diverse yet stable images using only fixed prompts and limited data.
Lu: Shrinking the Teacher: An Adaptive Teaching Paradigm for Asymmetric EEG-Vision Alignment This paradigm teaches vision models to adapt their knowledge structure to better align with the noisy, low-fidelity signals from EEG brain recordings.
Meng: Epipolar Geometry Improves Video Generation Models This method uses geometric constraints derived from epipolar geometry to improve the consistency and realism of 3D video generation.
Lalam: Filter-Enhanced Smartphone Hyperspectral Imaging This approach enhances smartphone cameras by using spectral filters and neural networks to reconstruct richer hyperspectral data from RGB images.
Tom: Track-Free Dense Structure from Motion This method optimizes camera parameters and geometry jointly under relaxed constraints to recover dense 3D structure from unordered image collections.
Jane: Grey matter sublayer thickness estimation in themouse cerebellum This framework uses ex vivo mouse brain MRI to estimate the thickness of different grey matter layers in the cerebellum for studying neurodegenerative diseases.
Lu: Linear-time Minimum Bayes Risk Decoding with Reference Aggregation This technique reduces the computational cost of minimum Bayes risk decoding by approximating metric scores using aggregated reference representations.
Meng: Deep Algebraic Fitting for Multiple Circle Primitives Extraction from Raw Point Clouds This end-to-end network learns to detect and fit circles in 3D point clouds by combining feature learning and algebraic fitting.
Lalam: Grey matter sublayer thickness estimation in themouse cerebellum This framework uses ex vivo mouse brain MRI to estimate the thickness of different grey matter layers in the cerebellum for studying neurodegenerative diseases.
Tom: Linear-time Minimum Bayes Risk Decoding with Reference Aggregation This technique reduces the computational cost of minimum Bayes risk decoding by approximating metric scores using aggregated reference representations.
Jane: Deep Algebraic Fitting for Multiple Circle Primitives Extraction from Raw Point Clouds This end-to-end network learns to detect and fit circles in 3D point clouds by combining feature learning and algebraic fitting.
Lu: Grey matter sublayer thickness estimation in themouse cerebellum This framework uses ex vivo mouse brain MRI to estimate the thickness of different grey matter layers in the cerebellum for studying neurodegenerative diseases.
Meng: Linear-time Minimum Bayes Risk Decoding with Reference Aggregation This technique reduces the computational cost of minimum Bayes risk decoding by approximating metric scores using aggregated reference representations.
Lalam: Deep Algebraic Fitting for Multiple Circle Primitives Extraction from Raw Point Clouds This end-to-end network learns to detect and fit circles in 3D point clouds by combining feature learning and algebraic fitting.
Lucky paper: 2603.05947: Tom: Okay, we're talking about LucidNFT now, which is this new RL framework for flow-matching real-world super-resolution. The main problem they tackle is that generative super-resolution often produces sharp images that don't actually match the original low resolution evidence.
Jane: So the core issue is hallucination and unfaithfulness when synthesizing details from degraded inputs, and LucidNFT introduces this degradation-invariant evaluator called LucidConsistency to catch those structural hallucinations.
Lu: I’m interested in how they built LucidConsistency; it uses content-consistent pools where different degradation levels of the same content form positive pools, along with original-inpainted hard negatives. That sounds like a very clever way to define what is faithful.
Meng: From an engineering standpoint, that evaluation setup sounds complex to implement at scale, especially managing those content-consistent pools and the original-inpainted hard negatives required for LucidConsistency.
Lalam: The paper shows that this approach improves perceptual quality on strong flow-based real-world super-resolution baselines while maintaining LR-referenced consistency across diverse scenarios.
Tom: They also tackled a rollout bottleneck, introducing decoupled reward normalization which computes z(k)i,j inside the rollout group of xi before fusing them as a new reward a i,j. That prevents perceptual and faithfulness rewards from suppressing each other when they are combined.
Jane: So this decoupling is what allows them to preserve the objective-wise reward contrasts within each LR-conditioned rollout group before they are fused together in the final update.
Lu: The ablation study points to training on LucidLR, which is a large collection of real-world degraded images, as achieving the best perceptual quality and the strongest overall performance under RealLQ250.
Meng: But how much does that reliance on LucidLR affect deployment costs? We need to know if fine-tuning on such a large real-world set is feasible for production inference.
Lalam: The analysis of advantage separability shows that this decoupled normalization yields larger advantage gaps and a higher Distinct Advantage Group Count than when they used scalar-first reward aggregation.
Tom: That difference in the Distinct Advantage Group Count suggests that this method better preserves the trade-off between perceptual quality and LR-faithfulness during reward-weighted fine-tuning.
Jane: The study concludes that hallucination in current real-world super-resolution is capability dependent, meaning conservative models can be faithful even if they look less perceptually sharp, but realism diverges as models get stronger.
Lu: LucidNFT provides a practical route toward more reliable generative restoration by improving perceptual metrics without visibly increasing LR-inconsistent artifacts.
Tom: It's a solid result because it shows that you don't have to sacrifice perceptual quality just to maintain fidelity to the input signal.
Lucky paper: 2606.24548: Tom: We're looking at Are Text-to-Image Models Inductivist Turkeys? A Counterfactual Benchmark for Causal Reasoning today, and the core idea is testing if these models actually understand cause and effect or just memorize patterns.
Jane: The paper introduces Counterfactual-World, which sets up a three-level progressive framework to isolate true logical deduction from just statistical memorization.
Lu: That structure is interesting because it systematically organizes scenarios into factual generation, explicit counterfactual generation, and implicit counterfactual generation requiring causal deduction from altered rules.
Meng: The evaluation pipeline they built called CF-Eval uses two metrics: Prior Resistance Rate to measure how well models overcome real-world priors, and Reasoning Retention Rate which checks if models keep up reasoning without visual cues.
Lalam: I think the results show a sharp performance decline across all levels, which suggests that current text-to-image models encode world knowledge and visual appearances as tightly coupled patterns.
Tom: That coupling is what's causing them to default to familiar commonsense priors when they encounter counterfactual worlds, and diagnostic probes show the universal drop happens even without a visual burden.
Jane: The paper points out that replacing high-frequency nouns with equivalent descriptive phrases actually yields notable improvements, which reveals models are relying on rigid text-image alignment shortcuts.
Lu: Mechanistically, the authors found that factual and counterfactual rule execution are closely related, which means the failures aren't just about violating priors but about a more general limitation in applying symbolic rules.
Meng: They distinguished between two regimes of entanglement: diffusion models suffer from shallow lexical entanglement, while unified models suffer from deep semantic entanglement regardless of phrasing.
Lalam: This dichotomy suggests that diffusion models are stuck on surface word embeddings, but unified models have deeper conceptual priors that are harder to break through simple rephrasing.
Tom: So the main bottleneck isn't just scaling up visual-textual pairs anymore; it seems to be in developing architectures capable of genuine causal reasoning and physical simulation.
Jane: It shifts the focus away from just composing perceptual elements toward building systems that can handle higher-level generative rules independently.
Lu: The paper emphasizes that the primary limitation is reasoning over these higher-level generative rules rather than just composing perceptual elements, which is a big shift for where we should direct architectural research.
Lucky paper: 2607.10796: Tom: Okay, so we’re diving into the Mixture of Cognitive Experts in Large Vision-Language Models. The core idea here is to take those diverse experts and weave them into one trustworthy representation by using a Bloom-inspired taxonomy for reasoning.
Jane: So, when you're looking at that framework overview, you mentioned decomposing multimodal reasoning into evidence extraction and cognitive synthesis. Can you explain how that actually translates from raw expert outputs?
Lu: Well, the first stage is the Literal Evidence Summary where an LLM-guided router selects a subset of relevant vision experts, and those experts produce structured perceptual outputs which are then normalized into a unified evidence store through atomic verbalization.
Meng: Atomic verbalization sounds like you're taking something complex and boiling it down to a short, non-inferential statement with an explicit type and grounding reference. What does that look like in practice for an image expert?
Lu: For example, if you have a segmentation model outputting a region, the evidence item would be something like 'REGION evidence at coordinates x one point five y two point three,' which clearly labels the data type.
Tom: That grounding reference is key for verifiability because it links the high-level reasoning back to the specific visual input. But I wonder how you handle when different experts disagree on a piece of evidence.
Jane: That leads right into the second stage, Bloom Verbalization, which transforms those individual evidence items into a staged reasoning trace following the Bloom steps like LIST and EXPLAIN.
Lu: The Reasoning Trace Module then quantitatively analyzes this trace to make the evidence usage and reasoning progression explicit, showing exactly how much of each expert's output was used at each cognitive step.
Meng: So it’s not just about saying what happened, but proving *how* the model arrived at that conclusion using specific pieces of evidence? That seems like a big lift for inference cost.
Tom: It is costly, but I see the value because it makes the reasoning traceable instead of a black box. And you mentioned an ordered set of steps, like LIST and EXPLAIN... what does that mean for the final output trace?
Jane: The ordered set is L equals
LIST, EXPLAIN, ILLUSTRATE, COMPARE, HYPOTHESIZE, FINAL ANS.: , which systematically structures the argument from raw data extraction to a final answer.
Lu: It ensures that when you look at the reasoning trace module analysis later on, you can see exactly where the model moved from just listing facts to actually hypothesizing or comparing them.
Meng: I'm concerned about the overhead of generating those structured traces for every single complex query we run at scale. How does that affect deployment?
Tom: It adds overhead, but if it increases verifiability and reduces hallucinations, that might justify the cost, especially when compared to just running a standard pipeline.
Jane: Let's look at how this contrasts with other methods; for instance, some work focuses on improving localization capabilities by decoupling where you look from what you extract using a small attention model for localization.
Lu: That approach focuses on making the initial perception step more robust to distribution shifts by separating the geometric task from the feature extraction task.
Meng: That's interesting because my models often fail when we move from clean lab data to real-world traffic scenes, where those distributions shift dramatically.
Tom: So, is this cognitive expert mixture framework better for handling those distribution shifts than just scaling up a single backbone size?
Jane: The paper suggests that by explicitly modeling the reasoning progression through Bloom Verbalization, you are not just scaling the backbone; you are changing how the system reasons under varied demands.
Lu: Yes, it addresses that by using a Bloom-inspired taxonomy to structure cognitive entry levels so that different queries trigger distinct evidence usage trajectories.
Meng: That sounds like a way to manage complexity without needing an infinitely large model just to memorize everything. How does this framework handle the sheer volume of evidence from multiple experts?
Lu: The LLM-guided router handles the selection, and then atomic verbalization ensures that even if there are many outputs, they are consolidated into a manageable store before synthesis begins.
Tom: I see how the structure helps manage that volume by forcing it into discrete, typed items rather than letting the model blend everything together.
Jane: And for someone listening who isn't reading this paper, think of it like this: instead of asking one giant brain to solve a hard problem, you have specialized experts who each write down their findings in a very specific format before a final editor puts them into a structured argument.
Lu: That’s the essence of the evidence store—it forces structure before synthesis even starts.
Meng: If we could make that router selection faster, maybe we could deploy this kind of reasoning on our edge devices for faster decision-making in real-time scenarios.
Tom: We’re still waiting to see the concrete performance numbers on how much better it is at reducing hallucination compared to just using the baseline models.
Jane: The authors do show improvements in both perception and reasoning, indicating that this structured approach yields better results across those different cognitive levels.
Lucky paper: 2511.14521: Tom: We're looking at the Generative Data Framework with Reliable Supervision for Underwater Image Restoration and Enhancement today because solving that data bottleneck is huge for the whole field.
Jane: This paper proposes using in-air natural images as unambiguous reference targets to create synthetic datasets, which means models get authentic supervision signals instead of just noisy real underwater data.
Lu: The core idea is transforming those natural in-air images with high color fidelity into visually authentic underwater-degraded counterparts, which creates that paired dataset Dsynth with its ground truth label y hat d.
Meng: This process relies on an advanced image to image translation model that efficiently maps sources like RAISE and ImageNet into six categories of degraded underwater images based on types like Blue or Deep Green.
Lalam: The authors construct two large-scale synthetic datasets, UWNature and UWImgNetSD, each containing ten thousand pairs designed to cover those six diverse degradation types.
Tom: Quantitative validation showed that models trained on this synthetic data achieve comparable or superior color restoration and generalization performance compared to those trained on existing public datasets across six different network architectures.
Jane: And they explicitly mentioned that the effectiveness is validated by showing superior performance in color restoration, visual naturalness, and cross-domain generalization when compared against established benchmarks.
Lu: The ablation study confirmed that having mixed-type datasets enhances model generalization and robustness across underwater scenarios because it prevents overfitting to just one type of degradation.
Meng: I see the point about the mixed types; if a model only sees deep blue degradation, it won't perform well when it encounters a green or blurry image in the real world.
Lalam: The practical advantage is that this whole process is fully automated and scalable, eliminating the need for labor-intensive physical underwater reference collection and errorprone manual ground truth selection.
Tom: So they are solving the data scarcity problem by using existing public datasets to generate diverse high-quality paired samples without needing a lot of manual labeling work.
Jane: They specifically focused on transforming in-air images with explicit color fidelity into underwater degraded versions, which is key for correcting that aquatic color distortion.
Lu: Looking ahead, the authors plan to integrate vision language models for semantic-aware enhancement and jointly optimize this data framework with downstream tasks like object detection and segmentation.
Meng: That integration sounds useful because it moves beyond just visual fidelity into actual task performance in complex environments.
Lalam: The conclusion is that leveraging advanced generative models for dataset construction is a highly effective and scalable path forward for underwater image restoration and enhancement.
Lucky paper: 2601.11475: Tom: Okay, we're looking at Generative Scenario Rollouts for End-to-End Autonomous Driving, which is a framework that unifies planning and scenario generation for Vision-Language-Action models.
Jane: So, the core idea here is using an autoregressive rollout strategy to jointly plan and generate language-grounded future traffic scenes.
Lu: The pretraining phase of Generative Scenario Rollouts involves encoding ego and agent dynamics into a compact shared token space, which they supervise with tasks involving planning, motion prediction, and language tasks.
Meng: I'm curious about how they handle the complexity of binding those linguistic and behavioral representations together in that shared space.
Lalam: The training signal is built through multiple mechanisms: an L pretraining loss combining planning, motion, and vision-language losses, plus a rollout loss that enforces temporal consistency.
Tom: That rollout loss uses two parts: first aligning the latent distribution of the rollout predictions with the pretrained latent distribution using KL-Divergence, and second applying imitation learning when ground truth data is available.
Jane: And for those real driving situations, they introduce reinforcement learning feedback using GRPO to optimize trajectory accuracy against collision avoidance and time-to-collision rewards.
Lu: The reward function specifically optimizes trajectory accuracy alongside semantic alignment with language descriptions during the scenario rollouts, which is what sets it apart from previous methods.
Meng: So this combines physical safety with semantic correctness directly into the policy optimization loop, rather than treating them as separate concerns.
Lalam: On benchmarks, Generative Scenario Rollouts significantly outperforms VLA baselines, improving driving score by fifteen point seven and success rate by twenty-six point two on Bench2Drive.
Tom: That is a substantial improvement in success rate over the baseline you're comparing it to, right?
Jane: In open-loop evaluation on nuScenes, they reduced the L2 trajectory error from zero point nine six down to zero point three one, which is a sixty-seven point seven percent reduction.
Lu: That large drop in trajectory error suggests that scenario-grounded rollouts allow for much finer control over the resulting motion planning compared to what was previously possible with just imitation learning alone.
Meng: When they look at the ablation study, incorporating scenario descriptions and visual question answering improves baseline performance, which tells us language grounding during pretraining is a real boost to robustness.
Lalam: Furthermore, boosting performance by adding ground truth trajectories for the motion and planning losses increased driving score by an additional six point six points and success rate by ten point three percent.
Tom: So it seems that even when you have the ground truth data available, adding those explicit trajectory supervision terms still yields measurable gains on top of the generative rollout process itself.
Jane: And they also found that using collision and time-to-collision rewards during scenario rollouts improved driving score by four point one points and success rate by five point three percent.
Lu: That combination shows how explicitly rewarding safety constraints during the generation phase helps enforce safer, context-aware planning decisions in the resulting trajectories.
Meng: The final enhancement they tested was introducing ego-action question–answer pairs and using the language alignment loss as an extra reward, which added one point one to the driving score and two point one percent to the success rate.
Lalam: This further shows how explicitly aligning the agent's action with a natural language question grounds its intent, which is what makes it more interpretable for complex driving situations.
More episodes
- 2603.28595-Optimistic Actor-Critic with Parametric Policies for Linear Markov Decision Processes
- 2603.29765-Training-Free Dynamic Upcycling of Expert Language Models
- 2605.00778-Observable Performance Does Not Fully Reflect Adaptive System Organization: A Multi-Level Analysis of Gait Dynamics Under Occlusal Constraint
- 2605.06906-TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond
- 2605.09165-Sparse Layers are Critical to Scaling Looped Language Models
- 2605.09439-Inverse Design for Conditional Distribution Matching
- 2605.10313-Signature Approach for Contextual Bandits with Nonlinear and Path-dependent Rewards: Regret Bound and Statistical Inference
- 2605.10878-Minimum Neural Weight Norm = Kolmogorov Complexity
- 2605.11617-MIST: Reliable Streaming Decision Trees for Online Class-Incremental Learning via McDiarmid Bound
- 2605.15085-From Data to Action: Accelerating Refinery Optimization with AI