Daily Summary for 2026-11-11
daily
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: It's the eleventh of November, twenty twenty-six, and this is the day's research.
Jane: 29 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 Read It Back.
The day's research: Tom: SpectraReward turns pretrained multimodal models into reward models for image generation reinforcement learning by measuring prompt recovery instead of preference labels.
Jane: That reuse of existing ability across various diffusion models is significant because it lets researchers improve generation performance with existing text alignment.
Lu: SelfSpectraReward creates a closed loop where the policy's understanding branch serves as the reward model for its generation branch.
Meng: VideoTreeSearch addresses errors in long video QA by framing it as an iterative self-correcting search over an adaptive temporal tree structure.
Lalam: The framework builds a non uniform tree using adaptive temporal segmentation so each node represents a semantically coherent segment.
Tom: An agent learns to navigate this tree using four discrete operations: zoom in, zoom out, shift, and answer for explicit backtracking.
Jane: This allows for recovery as learnable primitives instead of implicit behaviors during long video question answering tasks.
Lu: VideoTreeSearch shows gains over prior multi-turn cropping agents on benchmarks like CG-Bench and Haystack-Ego4D by substantial accuracy points.
Meng: A training pipeline generates multi step paths including deliberate detours into incorrect branches followed by recovery for supervised fine tuning.
Tom: That process builds upon the idea of adaptive temporal segmentation seen in other works like Searching Videos as Trees.
Jane: SierpinskiCam tackles geometry-guided approach limitations where the target camera moves away from the source trajectory using texture cues.
Lu: It augments guidance with Sierpinski dome texture cues that retain trackable features even under large viewpoint changes while using a reference video conditioning mechanism.
Meng: Video-FLAIR introduces a training framework that learns to select the right reasoning mode for each query using reinforcement learning during training.
Tom: It generates responses under all three modes for the same prompt and uses a composite reward based on correctness and cost to favor the most effective one over models like Qwen2.5-VL.
Jane: FogDrive provides a multi modal synthetic driving dataset specifically designed to test perception under graded fog conditions with physics based fog on both camera and LiDAR inputs.
Lu: This allows experiments to show that mixing fog densities improves 3D detection at no extra data cost and demonstrates radar can improve fusion at every density level.
Tom: Read It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation In this paper, we propose SpectraReward, a training-free reward function that turns pretrained MLLMs into off-the-shelf reward models for image generation reinforcement learning.
Jane: TryOnCrafter introduces a framework that uses a renderable 4D proxy to enable camera controllable video virtual try-on.
Lu: Adaptive Visual KV Cache Reuse for Efficient Multi-Turn Iterative Multimodal Reasoning proposes an adaptive visual KV cache reuse framework to reduce redundant processing in multi-turn visual reasoning.
Tom: TwinWeaver introduces a framework that serializes patient histories into text to build digital twins for cancer forecasting.
Jane: Tac-DINO develops Touch3D and GeoTAC to learn tactile features by aligning 3D vision and tactile data at the patch level.
Lu: WireSeg-32K presents WireSeg-32k, a synthetic dataset generated by a co-simulation pipeline for physically grounded wire segmentation.
Tom: RefineAny3D introduces a vision language model that corrects depth without predicting numerical values using visual alignment for monocular 3D detection.
Jane: PRIM introduces a causal meta-learning approach that uses Bayesian inference to identify the root causes of anomalies in complex systems.
Lu: SierpinskiCam proposes SierpinskiCam, which augments geometry guidance with texture cues to improve camera controllability in video retaking.
Tom: Video-FLAIR introduces Video-FLAIR, a framework that learns to select the appropriate reasoning mode for multimodal queries using reinforcement learning.
Jane: FogDrive presents FogDrive, a CARLA dataset with graded fog conditions for training autonomous driving perception models.
Lu: Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA introduces VideoTreeSearch, a framework that uses a self-correcting search over an adaptive temporal tree for grounded long video question answering.
Tom: GRAPE presents GRAPE, a unified architecture that uses graph attention to improve prototype based medical image classification.
Jane: RedPhys introduces RedPhys, an extensible platform for red-teaming physical world models and agents against adversarial attacks.
Lu: Counterfactual Reasoning for Robust Visual Question Answering proposes a novel training framework that uses counterfactual contrastive learning to improve out of distribution generalization in VQA models.
Tom: Human-in-the-Loop Signature Bootstrapping for UAV Hyperspectral PFM-1 Mine Detection studies how human inspection effort varies when detecting landmines using hyperspectral imaging on UAVs.
Jane: FOLIO introduces FOLIO, a training free focused semantic memory system to efficiently manage and retrieve information from continuously arriving video streams.
Lu: A Nearable Soft Mat Based on Distributed Optical Fiber Sensing for Physiological Monitoring describes a soft mat using distributed optical fibers for spatially resolved physiological monitoring of cardiorespiratory signals.
Tom: GRACE introduces GRACE, a framework that selectively erases sensitive concepts from text-to-image diffusion models while preserving generation fidelity.
Jane: LAFP proposes LAFP, which uses flow matching and interpolation to preserve the latent action structure during imitation learning.
Lu: Next Forcing introduces Next Forcing, a multi chunk prediction framework that accelerates training and inference for world action models.
Tom: Facial Expression Recognition in the Deep Learning Era is a systematic multi criteria review of methods models datasets performance challenges and future research directions.
Jane: CSFlow introduces CSFlow, a weighting scheme that connects human contrast sensitivity to flow matching to improve image generation quality.
Lu: Retrieve What's Missing proposes CovRAG, a depth based memory retrieval framework that maximizes coverage gain for consistent long video generation.
Tom: BIAS-ID introduces BIAS-ID, a transparent framework to analyze and quantify transformation biases in AI generated image detectors.
Jane: OmniMem proposes OmniMem, an explicit full range memory retrieval framework that uses sparse KV retrieval for long video generation.
Lu: Flow Mismatching introduces Flow Mismatching, an unsupervised anomaly detection method that identifies anomalies by comparing learned flow velocities to geometric paths.
Tom: A Causally Grounded Taxonomy for Image Degradation Robustness Evaluation proposes a causally grounded framework to organize and interpret image degradations based on their causal source and perceptual effect.
Jane: PointForward introduces PointForward, a feedforward reconstruction framework using point aligned representations for high fidelity driving scene reconstruction.
Lucky paper: 2607.11886: Tom: So we're looking at SpectraReward and SelfSpectraReward, which turn pretrained multimodal models into reward models for image generation reinforcement learning by measuring prompt recovery through a single image conditioned teacher forward pass.
Jane: That whole mechanism sounds like it reuses existing alignment knowledge from the MLLM without needing any external reward model or extra fine-tuning.
Lu: The reward function is defined as RM x y equals one over T times sum from t equals one to T minus one log p M of x t plus one x less than or equal to t y.
Tom: And the experimental validation covers two diffusion models, three reinforcement learning algorithms, nine reward MLLM backbones from four families spanning 4B to 235B parameters, and five out-of-distribution text-toimage benchmarks.
Jane: The analysis shows that token level likelihoods are sensitive to targeted semantic errors when comparing positive and negative image pairs for attribute mismatch and object identity mismatch.
Lu: SelfSpectraReward shows better improvements compared to SpectraReward with similar scale MLLM backbones, and it can match or surpass much larger external reward models.
Meng: Larger reward MLLMs aren't always better, but the self-reward can match or surpass those external models if the reward policy alignment is strong enough.
Tom: That suggests that distributional alignment between the reward model and the policy is a key factor for effective image generation reinforcement learning.
Jane: But what about where this approach stops working? The paper mentions that SpectraReward relies on the image conditioned prompt likelihood of pretrained MLLMs, which means its reward quality is bounded by the visual reasoning ability of the chosen backbone.
Lu: It also points out that because it computes the likelihood only over the input prompt, it might under-emphasize implicit visual implications that are not directly expressed in text.
Tom: And there's a practical detail on implementation, when using SpectraReward, we don't add any general prefix to ask the reward MLLM to describe the image by default.
Jane: The only exception they mention is for the InternVL3 point 5 family, where adding the prefix "Describe the image" gives more stable reward values.
Meng: I wonder how much effort it takes to integrate this into our current production pipeline compared to just using a standard preference label dataset for training.
Tom: Well, one of the broader impacts is that SpectraReward lowers the cost of constructing reward signals by reusing frozen pretrained MLLMs without preference annotation or reward-model fine-tuning.
Jane: That accessibility is something I think many researchers will find very useful when they have limited annotation resources for their image generation RL work.
Lu: However, the paper also warns that easier reinforcement learning optimization for image generation might amplify existing risks of text-toimage models.
Tom: Exactly, stronger prompt following can be misused to generate misleading visual content or imitate protected styles, so we absolutely need to combine this with standard safety filters and bias auditing.
Jane: What about the EOS token? The paper excludes it from the computation because its likelihood is often dominated by sequence termination rather than actual image-text alignment.
Lu: That's a good point because if you don't include that, you might be rewarding termination instead of true semantic alignment.
Meng: If we were deploying this at scale, I’d need to know how the computational cost scales when using the largest two hundred thirty-five billion parameter MLLMs mentioned in the validation.
Tom: The paper suggests that larger MLLMs might actually devote more capacity to advanced vision language reasoning and instruction following tasks, which is a potential trade-off.
Jane: So, while it's powerful for prompt following, it doesn't directly optimize criteria like aesthetics or safety on its own, which the authors suggest combining with specialized models.
Lu: That limitation means that to get truly high quality outputs, you still need those complementary reward signals for things like style or safety constraints.
Tom: So to summarize, SpectraReward and SelfSpectraReward significantly improve generation performance across various diffusion models and RL algorithms by reusing existing MLLM knowledge through prompt recovery likelihoods.
Jane: It’s a training-free way to create reward signals, which makes image generation reinforcement learning much more accessible for researchers without massive annotation budgets.
Lucky paper: 2601.20906: Tom: So we're looking at TwinWeaver, which is this framework that serializes longitudinal patient histories into text to build a pan cancer foundation model called Genie Digital Twin or GDT.
Jane: That sounds like it’s trying to solve the problem of modeling sparse, multimodal clinical time series by turning everything into text prompts for large language models.
Lu: The core idea is fine-tuning these large language models on procedurally generated text representations of patient journeys so they become these digital twins.
Meng: I'm curious about the training pipeline; how exactly are you splitting those patient trajectories at critical time points to create the input target pairs?
Tom: They preprocess longitudinal data, generate input target pairs by splitting patient trajectories at critical time points, and then convert those pairs into text prompts for fine-tuning.
Jane: And what is the structure of that text prompt? Is it just a simple sequence of events, or does it capture the continuous biomarkers too?
Lu: The input to the model is defined as Xp t equals sp Hp t where Hp t includes all historical measurements up to time point t.
Tom: So you're feeding the model both the specific measurement at that time and its history leading up to it.
Jane: And what about the target, Yp t? What exactly are those future predictions or event statuses they are trying to forecast?
Lu: The target Yp t is a set of future predictions for various tasks, including time-series forecasting and predicting a patient’s event status at a future time point.
Meng: For forecasting, what's the horizon? What is the forecast duration ∆t in this setup?
Tom: The second task is time-series forecasting where the target Yforecast p t forecast V prime comprises future values for sampled variables V prime up to a forecasting horizon t plus delta t.
Jane: And for event prediction, what are the labels they are using there? Are those just binary outcomes like occurred or not occurred?
Lu: The set of event labels is Y equals occurred, not occurred, censored. So it covers whether an event happened, didn't happen, or if it was censored.
Tom: I see. They use a loss function L theta equals minus X N i equals k log q theta x i x one:i minus one where q theta is the language model parameterized by theta.
Jane: That looks like a standard language modeling setup, but you’re using it to predict clinical outcomes, which is interesting because of the downstream tasks.
Lu: The final ranking score for an event E is taken from the occurred class probability using log likelihoods derived from softmax normalization.
Meng: When you look at performance, what are the baseline comparisons you used when you reported that GDT achieves a median Mean Absolute Scaled Error of zero point eight seven?
Tom: You compared it to a baseline of zero point nine seven for the strongest time-series baseline and achieved an average concordance index of zero point seven zero three across survival, progression, and therapy switching tasks.
Jane: A C-index of one point zero is perfect discrimination, so going from zero point six six two to zero point seven zero three sounds like a pretty solid improvement in risk stratification capability for those clinical outcomes.
Lu: The framework also enables an interpretable clinical reasoning extension by fine-tuning LLMs on a synthetic corpus and then aligning that with group relative policy optimization or GRPO.
Tom: That alignment step is what gives you that transparent, biologically grounded context for hypothesis generation, right?
Jane: It sounds like they are trying to make the LLM's reasoning chains correspond to clinically relevant features and genetic contexts.
Meng: Does this extension work well when you test it against out of distribution clinical trials where historical data is sparse?
Lu: In those out of distribution clinical trial tasks, the fine-tuned GDT shows strong performance, achieving the lowest MASE in blood biomarker forecasting at zero point eight eight three and highest C-Indexes for survival and progression events.
Tom: That low MASE number is good because it means the model provides superior forecasts compared to just carrying forward the last known value.
Jane: You mentioned they also show performance enhancement when incorporating richer patient histories for survival predictions; what kind of additional data sources are they incorporating there?
Lu: The paper mentions that the reasoning extension provides a calculated trade-off for clinical utility by providing transparent, biologically grounded context for hypothesis generation.
Meng: I'm wondering about the synthetic example; does it actually generate reasoning chains that look clinically useful, or is it just noise formatted nicely?
Tom: A synthetic patient example demonstrates how the framework handles input and generates structured reasoning chains, including patient summary and mechanistic analysis.
Jane: So they are showing that the model can produce explanations for a patient's response in clinical practice based on that structured chain.
Lu: That capability allows for hypothesis generation by providing different explanations for a patient’s response in clinical practice based on their history.
Lucky paper: 2609.03102: Tom: We’re moving into paper number three of five today, and this one is WireSeg-32K, a physics grounded synthetic dataset for wire instance segmentation.
Jane: It's interesting because it’s combining thirty-two thousand RGB images with real world test set annotations.
Lu: The authors developed DeformX, which is a co-simulation pipeline coupling Cosserat rod dynamics with Isaac Sim rendering to generate physically plausible wire shapes and scenes.
Tom: So they are using Cosserat rod dynamics to model each deformable linear object as a continuous representation with a centerline and an orthonormal material frame.
Jane: That continuous representation lets them capture stretching, shearing, bending, and twisting while allowing for physically interpretable control over things like gravity driven sagging.
Lu: They also implemented free form mesh contact which computes closest point distances between rod nodes and mesh surfaces to allow wires to drape over structures or wrap around obstacles.
Tom: The dataset construction involves assembling scene templates from public assets and varying parameters like wire count, lengths, radii, and material properties for generation.
Jane: They categorized the scenarios into three types: wire on plane, flying wires emphasizing gravity driven dangling behavior, and data center settings with cluttered racklike environments.
Lu: The difficulty splits are determined by Segment Anything Model three performance on generated images using AP at seventy-five scores to label samples as easy or medium.
Tom: The evaluation shows that fine tuning Segment Anything Model three using LoRA adapters consistently improves performance over the off the shelf baseline across all subsets, including real images.
Jane: That suggests the diversity and physical realism of WireSeg-32K provide useful supervision for challenging wire perception even without real image training.
Lu: They also compared it to prior DLO datasets and found that WireSeg-32k is the only one listed that possesses physics realism, instance labels, CAD support, and visually grounded scenes all at once.
Tom: The results indicate a measurable synthetic to real transfer value even without real image training when fine tuning SAM3 on WireSeg-32k alone.
Jane: How does this physical modeling actually translate into what an agent learns when it tries to segment wires in the real world?
Lu: DeformX models each deformable linear object as a Cosserat rod parameterized by a centerline and an orthonormal material frame along arc length, which captures complex dynamics.
Tom: That continuous representation allows for physically interpretable control over things like bending stiffness and torsion, which is something you don't get from just looking at pixels.
Jane: I wonder if that level of physical detail is necessary or if simpler geometric cues would suffice for a standard segmentation model.
Lu: The authors conclude that this combination of Cosserat rod dynamics, free form mesh contact, lightweight co simulation, and mesh skinned rendering solves the problem of segmenting deformable linear objects.
Tom: So they are basically using physics simulation to create perfect training data for a deep learning model to learn how to see wires.
Jane: That's a pretty concrete application for this kind of detailed synthetic data.
Meng: From an engineering standpoint, how does the pipeline handle the multi rate interface between the rod engine and Isaac Sim?
Lu: They use a lightweight multi rate interface allowing fine substeps for the DLO solver while Isaac Sim updates rigid bodies at a coarser rate, which keeps things efficient.
Tom: That sounds like they managed to keep the simulation fast enough for practical data generation rather than just getting high fidelity with massive computation time.
Jane: It’s important that they have those specific difficulty splits based on AP at seventy-five because if you just train on easy scenes, you won't improve performance on harder real world cases.
Meng: I'm curious about the cost of running this whole DeformX pipeline; is it something that breaks at scale when generating thousands of configurations?
Lu: The complexity comes from simulating the DLO dynamics and the photorealistic rendering, but they designed it to be lightweight for this specific purpose.
Tom: So, if we take this finding about synthetic data transfer, what does that imply for deploying segmentation models in environments where real labeled data is scarce?
Jane: It suggests that investing in high fidelity physics simulation can provide a meaningful supervisory signal even when you are training on limited real world examples.
Meng: If we were to deploy this, the main question would be how much computational overhead is required for the inference part of using SAM3 with LoRA adapters on these generated images.
Lu: The paper shows measurable synthetic to real transfer value, implying that even with a small fine tuning process, you get gains on harder scenes.
Tom: So it's not just about getting more data; it's about the *quality* of the physics in that data helping the model generalize better.
Jane: And looking at how they compared it to other datasets, WireSeg-32K achieved physics realism, instance label, CAD support, and visually grounded scenes simultaneously.
Lu: That combination across all four metrics is what sets it apart from previous DLO datasets like those mentioned in Table two.
Tom: It’s a solid benchmark if you want to study sim to real transfer specifically for thin deformable objects.
Lucky paper: 2608.09147: Tom: Okay, we're moving into paper number four of five which is RefineAnythree dee. This paper addresses the lack of object level precision in current depth foundation models for monocular three dee detection by treating depth refinement as a stand alone task.
Jane: So, the core idea here is that instead of trying to get the foundation model to predict a precise numerical value for depth, they are reframing it as a visual alignment problem in image space.
Lu: Exactly. The key insight is that when you project the candidate three dee box onto the image plane, what matters is whether that projected box tightly encloses the target object.
Tom: They make this concrete by extending the vision language model's vocabulary with six new special tokens for direction and magnitude, which simplifies decoding and supervision.
Jane: The direction tokens encode qualitative outcomes like depth closer or depth farther based on how small or large the projected box is compared to the target object's size.
Meng: I’m curious about those magnitude tokens; they are tied to the object’s own characteristic size, which they calculate as one over three of its width, height, and length.
Tom: That means step sizes are relative to the object scale rather than being fixed numbers, which makes sense for generalization across novel categories.
Jane: They train in two stages; stage one just stabilizes those action token embeddings by training them separately before they jointly fine-tune the whole model in stage two.
Lu: Stage two is where they jointly fine tune the language model backbone and those action tokens while freezing the vision encoder to keep those general visual priors intact for novel scenes.
Tom: In terms of performance gains, applying RefineAnythree dee on top of DetAnythree dee lifts APthree dee by four point three five overall through depth refinement alone.
Jane: That improvement comes from refining pseudo labels for auto-labeling, lifting the downstream detector by two point one nine on easy, two point three seven on moderate, and zero point seven two on hard cases compared to ground truth trained accuracy.
Meng: The overhead is pretty heavy though; they use eight NVIDIA H100 GPUs with bfloat16 precision and stage two takes about ninety hours per configuration for training.
Tom: And inference adds about two point eight seconds per image because the mean number of vision language model calls is around five hundred twenty milliseconds.
Jane: The limitation they mention is that this visual signature cue weakens when objects are heavily occluded or truncated, which makes refinement less reliable under tight three dimensional intersection over union thresholds.
Lu: They also state a scope condition: the method depends on the upstream detector providing a reasonable initial three dee box, which is something we should keep in mind for deployment.
Tom: So, how does this change things for someone who only listens to the show? It means that for autonomous driving perception, more accurate object localization translates directly into better trajectory prediction and collision avoidance.
Jane: That’s right; more accurate object localization helps especially in the long tail of open vocabulary where current detectors are weakest.
Meng: I wonder if we could deploy this faster if we just used a simpler, fixed set of direction tokens instead of all six action tokens they introduced.
Tom: But they deliberately keep the chain-of-thought prefix that the model produces before emitting an action token because it acts as a language anchored intermediate signal.
Jane: That CoT essentially lets the model articulate which object it is looking at and what visual cue it is reading before committing to an action.
Lu: It’s interesting how they built that training data by synthetically perturbing each ground truth box along the camera frame depth axis to generate about three million training samples spanning all outcomes.
Meng: That synthetic data generation pipeline sounds intensive, but it seems necessary for covering all those action token outcomes.
Tom: RefineAnythree dee is a plug and play design because practitioners can integrate it on top of existing detectors without retraining the underlying perception stack.
Lucky paper: 2605.08786: Tom: Alright, we're looking at PRIM now. It frames root cause analysis as a Bayesian inference task over a prior of causal models because traditional RCA struggles with error propagation and structural knowledge.
Jane: So, the core idea is they are marginalizing out structural uncertainty to implicitly find changes in the data-generating mechanism between normal and anomalous periods without explicit statistical testing.
Lu: They achieve this using a Model Averaged Causal Estimation transformer neural process that jointly attends over observational and anomalous samples along with the structural dependencies of nodes, enabling zero shot inference in seventeen milliseconds for systems with up to one hundred variables.
Meng: Seventeen milliseconds is fast, but what about scalability when we move beyond one hundred variables? How does PRIM handle inputs that are much larger than that?
Lalam: The paper states that Scalability is assessed across randomly sampled DAGs of varying size, showing that Recall at one degrades more slowly with increasing graph size than any other method.
Tom: That's good, but the limitations section mentions identifiability rests on assumptions like acyclic SCM and causal sufficiency, and a misspecified prior can degrade accuracy by failing to capture the true root cause distribution.
Jane: So if the prior doesn't capture things right, we might get inaccurate results even though they aren't explicitly fitting a graph at test time.
Lu: And I should mention that PRIM cannot currently exploit a ground truth graph when one is available, which is something researchers often look for in causal modeling.
Tom: It does confirm that joint information about the target variable T and the anomaly can be informative in three node confounder and mediator topologies, multi root cause settings, and larger graphs.
Jane: That sounds like a lot of complexity to manage all at once when you're trying to diagnose a fault.
Meng: And for practical deployment, it takes about three minutes on a GPU for lightweight finetuning on a domain specific simulator to get meaningful gains over the zero shot baseline.
Tom: The inference time remains constant at seventeen milliseconds on an A100 GPU, which makes it very practical for real time fault diagnosis systems.
Jane: That level of efficiency is impressive when you think about running these kinds of analyses constantly in a production environment.
Lu: On the realistic benchmarks PetShop and CausRCA, PRIM is competitive with both graph free and graph based RCA methods when the prior captures the target system well.
Tom: The fine tuned variant, PRIM FT, is actually reported as the best graph not given method on both full and subsystem evaluations on CausRCA.
Jane: That suggests that for systems where we don't know the exact causal structure, this approach has a strong chance of working well.
Lu: Performance really depends on the mechanism family used; neural network mechanisms produce more distinctive distribution shifts than Gaussian Process mechanisms in this setup.
Meng: So, if I were building a system today and didn't have a graph but wanted to do root cause analysis, would I lean towards using the neural network mechanism over something like a Gaussian Process?
Tom: Based on what they measured, yes, the neural network mechanisms seem to produce more distinctive shifts.
Jane: So it seems like this approach is really trying to bridge that gap between needing a perfect causal map and not having one at all.
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