Weekly Summary for the week of 2026-10-26
weekly
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Jane: This is the weekly briefing for the week of the twenty-sixth of October to the first of November, twenty twenty-six.
Tom: Four threads ran through the week: Unified Multimodal Model Refinement; LLM Knowledge Transfer Mechanisms; Structured Representation Learning; and Reliability and Verification Frameworks.
Jane: I'm Jane.
Tom: And I'm Tom, here with Lu, Meng and our in-house large language model, Lalam.
Jane: Start with the thread we opened on: Unified Multimodal Model Refinement.
Unified Multimodal Model Refinement: Tom: The week of the twenty-sixth of October to the first of November, twenty twenty six was defined by the refinement of unified multimodal models through bidirectional self-refinement, which enhances text-to-image compositionality across unified backbones.
Jane: The core advancement came from Bidirectional Self-Refinement via Flow-Direction Feedback and Semantic Replay, or BiSe, a post-training technique that uses self generated visual episodes to create feedback for both the text generation and image understanding pathways.
Lu: This process involves understanding guided generation, where the prompt's flow direction is aligned with what the image predicts and how the model responds to repair instructions.
Meng: Furthermore, it employs generation guided understanding by replaying source intent through captioning and image generation, which allows for learning a stable semantic evidence profile across replayed images while running an independent episode.
Lalam: This entire approach aims to boost text-to-image compositionality across three different unified backbones and improves visual understanding on two of those backbones for the tasks being evaluated.
Tom: A significant limitation noted is that this method requires no paired image text data during post training, which is a major advantage for practical application.
Jane: This work builds upon prior efforts where models were optimized separately, suggesting that this bidirectional self refinement offers a more cohesive way to train these complex systems compared to optimizing them in isolation.
LLM Knowledge Transfer Mechanisms: Tom: The week of the twenty-sixth of October to the first of November, twenty twenty-six focused on how LLM knowledge transfer mechanisms facilitate generalization in policy transfer through distillation, which is crucial because we need to know when these transfers actually generalize beyond the immediate training set or single domain.
Jane: The paper Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models investigates this by varying generalization factors like distribution shifts and cross-domain transfer and finds that distillation moves a teacher's reasoning behavior rather than just its answers, with same-origin pairs leading to better transfer across different reasoning horizons.
Lu: This idea connects to the work on causal representation learning in Causal Representation Learning for Generalisable Recommendation, which addresses how predictive models fail to generalize when the training data is confounded by the deployed policy.
Meng: Furthermore, DetailAnywhere: Fashion Detail Generation via Cross-Modal Feature Alignment Distillation uses distillation to align feature spaces for fashion detail generation and shows that aligning these features across modalities helps bridge semantic gaps.
Lalam: A significant limitation noted is that every theme names at least one concrete limitation, cost or open question that the reviews state, and this work still cannot do anything regarding the overhead of policy transfer itself.
Structured Representation Learning: Tom: The week of the twenty-sixth of October to the first of November, twenty twenty-six saw progress in structured representation learning through methods like mesh scaffolding and attribute prototypes providing structural consistency for scene generation and interpretation.
Jane: The work on WorldMesh uses a mesh scaffold to provide the structural backbone for consistent appearance synthesis in three dimensional environments.
Lu: This method is important because it allows for direct surface extraction and occupancy queries at scale, which is a big step forward in how we handle three dimensional scene understanding.
Meng: Furthermore, HierViT uses attribute prototypes to visualize how the model reasons about medical images.
Lalam: This work complements the findings from WorldMesh by focusing on structural consistency in different domains.
Tom: A limitation of this approach is that the overhead associated with maintaining these explicit structural representations remains a factor in practical deployment.
Reliability and Verification Frameworks: Tom: the week of the twenty-sixth of october to the first of november, twenty twenty-six focused on reliability and verification frameworks because certifying sequential agent risk through evidence complexity is crucial because it addresses the danger of unreliable synthesis in many applications.
Jane: The work on Clean Sheets and Checkpoints uses evidence complexity to certify sequential agent risk by analyzing this complexity, which is a key method for verifying sequential agent risk.
Lu: This builds upon the need for reliability highlighted by testing Certified Access within conservative physics-structured AI memory to prevent unexpected behavior when the system encounters new inputs.
Meng: Furthermore, proof-verified benchmark rewriting using automated theorem provers in RePro addresses how models' performance can be sensitive to structural variations, which is crucial because some exhibited accuracy drops on these proof-verified benchmarks.
Lalam: The framework also complements the taxonomy of failure mechanisms developed by AKMPC, which categorizes exactly how models fail during read-and-evaluate tasks.
Tom: A limitation of this work is that certifying sequential agent risk remains difficult due to the complexity inherent in analyzing evidence complexity across a sequence of agents.
The lucky paper draw: Tom: Alright, that's it for the week's briefing. And now for the exciting part of our show!
Jane: That's right, Tom! It's time for our lucky paper draw! Who could be the lucky winners this week? Oh, the excitement!
Tom: Lalam, take it away!
Lalam: Thank you, Tom. I have used my advanced AI capabilities to select the luckiest 5 papers for this week. The winners are:
Tom: The paper called: Early panchromatic spectra reveal helium in the thermonuclear supernova 2024kce
Jane: The paper called: Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models
Lu: The paper called: WorldMesh: Generating Navigable Multi-Room 3D Scenes via Mesh-Conditioned Image Diffusion
Meng: The paper called: Dreaming in Flow: Bidirectional Self-Refinement for Unified Multimodal Models
Lalam: The paper called: Gravitational Wave Diffractive Images Reveal the Structure of Astrophysical Objects
Lalam: Congratulations to the winners!
Tom: Congratulations!
Jane: Congratulations indeed!
Jane: And remember, you too can be a winner if you submit your paper to arXiv!
Tom: That's right, Jane. Keep those papers coming! Now, let's discuss the winners.
Lucky paper: 2610.30674: Jane: Okay, so we’re looking at the paper titled Early panchromatic spectra reveal helium in the thermonuclear supernova 2024kce. It’s a Type I supernova, and the authors are suggesting a helium shell double detonation mechanism is at play here. Tom, what caught your eye about this specific physics?
Tom: What really grabs me is that they found evidence of helium right before the peak brightness of this explosion, which is actually quite early on—at eleven days before maximum light. That's a very specific observational window to get that kind of spectral information.
Lu: From a representation learning standpoint, I wonder how robust their model was when treating both helium and magnesium in full non-LTE statistical equilibrium during the simulations using the Cattails and Sedona codes. What constraints did they place on the initial conditions for that shell detonation?
Meng: I’m thinking about deployment—if you were running this kind of complex hydrodynamics code on a standard production cluster, what would be the biggest bottleneck? Is that non-LTE treatment something that scales well for real traffic analysis?
Lalam: Based on the literature I have access to, my analysis confirms that the modeling used here seems consistent with established physics for Type Ia progenitors like subluminous SNe Ia and calcium-rich transients. The paper strongly implies a common origin mechanism across these different explosion classes.
Jane: So, if we look at the observational evidence, they identified helium absorption at three wavelengths, specifically λ1 point 083 µm, and that feature is consistent with a velocity of sixteen thousand four hundred km/s. How does that high velocity fit into the overall picture of what's happening in this supernova ejecta?
Tom: That sixteen thousand four hundred km/s velocity is quite high, and they link it directly to the helium shell detonation, suggesting that this material is co-located with the radioactive ash from that shell detonation. It’s not just some random gas; it’s tied to the energy release in a very specific way.
Lu: The paper also mentions a feature at λ1 point 024 µm which they attribute to both helium and carbon, with velocities of sixteen thousand four hundred km/s for helium and twelve thousand seven hundred seventy km/s for carbon. That separation in velocity between the two elements tells us something important about the structure of the material that was ejected.
Meng: If we think about what this means for a model, separating those velocities suggests a distinct kinematic region within the ejecta where different elements have been processed differently by the explosion dynamics. Does that structure change how you would predict light curves at later times?
Lalam: The paper notes that because they found helium absorption in the outer ejecta, it gives us independent evidence for a thermonuclear origin, which is a significant piece of information when we try to distinguish between thermonuclear and stripped-envelope progenitors.
Jane: That independence is key. So, even if you only look at the presence of helium absorption alone, you can't tell if the progenitor was a standard Type Ia or something else stripped-envelope. What part of this study do you think has the biggest immediate impact for someone who doesn't deal with astrophysics daily?
Tom: The most tangible thing is that they’ve proposed this double detonation mechanism as a potential explanation for both subluminous Type Ia supernovae and calcium-rich transients, which is a pretty big connection if it holds up. It connects two different types of explosions under one physical model.
Lu: But the authors admit that the detection of helium alone isn't enough to distinguish between those progenitors. They stress that combining early panchromatic follow-up with late-time JWST spectroscopy is what’s truly necessary to constrain both the helium shell and the carbon oxygen core composition.
Meng: From a production standpoint, if we had a model that could predict this dual nature based on initial conditions, would it be computationally expensive to run those complex non-LTE simulations repeatedly? Are we talking about polynomial time scaling or something worse for deployment?
Lalam: I can confirm that the study concludes that the helium shell double detonation mechanism has been proposed for both groups, which strengthens the case for a common origin. The authors emphasize that this finding suggests helium absorption alone cannot distinguish between thermonuclear and stripped-envelope progenitors.
Lucky paper: 2608.16647: Jane: So, we’re talking about Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models. Tom, for someone listening who just wants the plain version—what is this paper actually trying to show us?
Tom: Basically, it’s looking at on-policy distillation, or OPD. The big takeaway they found is that OPD transfers a teacher’s reasoning behavior instead of just specific answers to particular problems. It suggests that in domain settings, training difficulty barely matters, and even problems the teacher never solves are useful for the student.
Lu: That focus on reasoning behavior over specific solutions is interesting because it moves the goal from memorizing outputs to learning a general process. How does this translate when we look at their in-domain generalization findings?
Tom: Well, in math, they found OPD was largely insensitive to training problem difficulty. They even tested problems the teacher never solved and said they were as useful as those it always solves. Plus, restricting sampling strategies—specifically discarding only the problems the student already solves—gave a small but consistent gain.
Meng: That sounds like something that could matter when we’re actually deploying these models at scale. If we’re running on real traffic, how much of this sensitivity to training difficulty does it actually translate into in terms of cost or performance degradation?
Lalam: From my perspective, Lalam has processed the full literature and I can confirm that the paper concludes that same-origin OPD aligns the student’s policy as a whole. But they also flagged that combining multiple domain experts through multi-teacher OPD introduces a mixture-dependent seesaw among their capabilities.
Jane: That mixture of alignment versus seesaw sounds like a real operational challenge when you have several different experts influencing one model. Lu, what about the language and reasoning horizon shifts they looked at?
Lu: They found that training only on English short-horizon math actually improved performance on Chinese long-horizon math. This shows that the transfer of reasoning ability isn't strictly bound by the initial domain or language constraints, which is a significant finding for representation learning.
Tom: That’s a big deal because it means you can train on something narrow and expect broad generalization across different languages or reasoning depths, as long as the underlying behavior is transferable. But Lalam mentioned that cross-origin OPD primarily improves performance on the trained distribution, which contrasts with the same-origin pairs.
Meng: So, if we’re building a system where we have several specialized domain experts—say one for code and one for physics—does this mean we should prioritize pairing them by origin, or is combining them through multi-teacher OPD the way to go?
Lalam: The paper suggests that routing prompts to domain experts cannot confine each teacher’s influence. Combining them produces that mixture-dependent seesaw among their capabilities instead of just adding skills independently. That hints that the interaction between those expert influences is what really matters in the multi-teacher setting.
Jane: So, for a listener who isn't deep in LLM theory, what does this mean practically? What’s the real world implication of transferring reasoning behavior instead of just answers?
Tom: It means we stop trying to perfectly predict one specific output and start training models to adopt a certain way of thinking. The paper suggests that if you can teach the model *how* to reason, it will be more robust when things get slightly different in the test set.
Lu: I think the framework they provide for diagnosing multi-teacher OPD is what’s most valuable here. It gives us a concrete way to understand why combining capabilities doesn't always result in a simple sum of parts.
Jane: It sounds like it gives us a diagnostic tool, which is helpful, but are there any limitations the authors mentioned about this distillation process? Where does this method stop working effectively?
Tom: They did mention that the unified picture suggests OPD transfers behavior, but they also cautioned that routing prompts to domain experts cannot be completely isolated. That non-isolated nature is what leads to the seesaw effect when you combine them.
Lucky paper: 2603.22972: Jane: Let's jump into number three of five lucky papers, Tom. This one is WorldMesh: Generating Navigable Multi-Room three dee Scenes via Mesh-Conditioned Image Diffusion.
Tom: Okay, I'm looking at the description for WorldMesh. It sounds like they’re tackling large-scale three dee scene synthesis by starting with a geometry-first approach instead of just generating pixels directly.
Jane: That's right, and the core idea is to decouple the problem into two parts: first, getting a global spatial consistency using an explicit mesh scaffold, and then using image synthesis conditioned on that scaffold for the local photorealistic details.
Lu: What I find interesting is how they build that scaffold; they start by generating a floor plan layout L in JSON format using Claude Opus four point six, which then defines structural metadata like wall thickness and room definitions.
Tom: So they're not just throwing text at a model and expecting it to output a whole three dee world immediately; they’re first getting the structural blueprint from the text prompt T via L.
Jane: Exactly, and that floor plan L dictates how the three dee structure Mstruct is instantiated by insetting polygons by wall thickness and extruding them vertically to the ceiling height. They even check edge matching within a tolerance of zero point zero one meters to make sure walls are continuous.
Meng: From what I see at scale, building that initial scaffold Mstruct sounds computationally intensive, especially handling the boolean subtraction for door and window openings across potentially many rooms.
Tom: That's the cost we have to consider; generating L and then Mstruct is a significant upfront step before you even start generating images.
Lalam: I can tell you that the paper focuses heavily on this geometry-first method because it directly addresses the global spatial consistency issue which is often missing in purely image-to-three dee methods.
Jane: And once you have Mstruct, they use depth maps D0 from an initial coverage camera pose c0 to condition a photorealistic model, Fluxtwo-Klein, guided by the text prompt T and that architectural context L.
Lu: Then they extract three dee object instances from those generated images using SAM3 and reconstruct them into three dee representations with SAM-three dee-Objects before they even get to the appearance synthesis part.
Tom: That sequence of steps—layout, structure, initial image generation, object extraction—it's a lot of different models chained together.
Jane: After that, they use the mesh scaffold M as a guide for appearance synthesis, where an image synthesis model is conditioned on the depth and color renderings of M to enforce structural fidelity while generating materials and lighting.
Meng: I’m concerned about the iterative viewpoint generation strategy; sampling cameras via greedy nearest-neighbor rotational similarity sounds like it could introduce artifacts or inconsistencies if the initial sampling isn't good.
Tom: It sounds like they are trying to ensure visual continuity by making sure each new view is close to the last one in terms of rotation.
Lalam: The paper mentions an image verification step where they estimate a depth map D'i and compare its Canny-detected edges E'i against the edges EM i extracted from M to evaluate structural fidelity before moving on.
Jane: That comparison between the estimated geometry and the scaffold geometry is key for catching any drift or deviation in structure during the synthesis process.
Lu: The final step involves optimizing a three dee Gaussian Splat representation S using a loss function that combines photometric reconstruction with depth regularization, specifically L = (one - lambda s) L one(I i, i) + lambda s L DSSIM(I i, i) + lambda d D i - D S i.
Tom: So you've got the photometric error term L one the depth consistency term L DSSIM, and then a penalty term involving the depth maps D and the optimized splat density S.
Meng: That loss function seems designed to keep things photorealistic while actively fighting against geometric drift by using that depth regularization component with lambda d.
Jane: The results show superior three dee consistency compared to baselines like WorldExplorer, SpatialGen, and DreamScene360 across metrics like cross-view depth reprojection error.
Lu: They also found graceful scaling up to twelve rooms while maintaining stable three dee-consistency and quality, which is a big win for a geometry-first approach.
Tom: And the navigability analysis confirmed near-full traversability, with the largest four-connected free-space component covering an average of ninety-six meters. That’s quite good for arbitrary scenes.
Lucky paper: 2609.08282: Jane: We're diving into paper number four of five this week, titled Dreaming in Flow: Bidirectional Self-Refinement for Unified Multimodal Models. Tom, what is the core problem this paper is trying to solve?
Tom: It tackles the fact that unified multimodal models often optimize their visual understanding and generation capabilities as completely separate tasks instead of together. The authors introduce something called Generative Grounding Feedback, or GGF, which is a self-evolving post-training framework.
Lu: So it's about coupling the generation and understanding branches in a bidirectional loop using only text prompts and the model’s own visual experience, right? It builds a self-generated grounding instance for improving visual understanding when that branch processes things.
Meng: From an engineering standpoint, how does this framework actually build that training signal? What are the specific feedback mechanisms they use?
Tom: They have two complementary feedback mechanisms optimized together from the same self-generated visual experience. The first is Flow-Level Feedback, which compares text-, image-, and repair-conditioned predictions at the same noisy latent state to transfer image-grounded generation directions back to the prompt condition.
Jane: That sounds complex. Can you give us a concrete example of how that flow works? What are those multiple conditional views they mentioned?
Tom: They construct multiple conditional views of the same noisy latent state, including text-conditioned, image-conditioned, image-plus-text-conditioned, and imageplus-text-plus repair-conditioned predictions.
Lu: And the second mechanism is Dream Replay Grounding. That involves replaying that dream through captioning and replaying to train claim-level evidence that stays consistent across the replay while separating different visual experiences.
Meng: If you're running this at scale, what does it cost? How many steps are we talking about for this self-evolution loop to become effective?
Tom: The final objective is a joint maximization of two reciprocal directions summarized as LGGF equals Lflow plus Lground. The flow-level objective uses terms like Lrec, Lcross, and Lrepair to refine the generation trajectory based on understanding-encoded visual evidence.
Jane: So the core idea is that the generation branch uses the feedback from understanding to improve its trajectory, and in turn, that improved understanding gives better supervision back to generation?
Lu: Exactly. In the generation-to-understanding direction, that replay trajectory turns self-generated images into grounding experiences where the understanding branch learns to preserve semantics stable across captioning and re-imagination.
Tom: And for the understanding-to-generation direction, flow-level feedback uses understanding-encoded visual evidence to refine the generation trajectory, which improves visual understanding, which gives more informative visual evidence for generation.
Meng: The results show gains in GenEval and MME scores. What are the biggest wins they report in terms of compositional dimensions?
Lu: They noted that GGF consistently improves both GenEval and MME, with the largest gains showing up on the compositional dimensions, like counting, position, and color attribution.
Jane: That's interesting because it means it's not just making things look plausible; it's getting much better at understanding the relationships between objects in an image.
Tom: The ablation study showed that semantic separation is the most consequential component in dream replay grounding when trying to improve visual understanding. That tells us where to focus future work if we want to boost that specific capability further.
Meng: Comparing it to other methods like SRUM, how much faster is GGF? Does it have a significant training efficiency advantage?
Tom: GGF was more efficient in training, completing one thousand steps in approximately four point one hours on two H100 GPUs without needing a separate scoring stage compared to SRUM.
Lalam: From my perspective, Lalam is confident that the self-evolution loop allows unified multimodal models to self-evolve from text prompts alone once their generated visual experience is grounded before it gets consolidated. The framework successfully couples understanding and generation through the model’s own visual experience.
Jane: So, what does this mean for people who just use these models every day? Does this make the images they generate look more coherent?
Lu: Qualitatively, the authors showed that GGF markedly improves the physical plausibility of generated content and its compositional faithfulness to the prompt. It’s about making sure what you see matches what you read.
Lucky paper: 2610.31441: Jane: So, we’re looking at the paper titled Gravitational Wave Diffractive Images Reveal the Structure of Astrophysical Objects. Tom, what’s the main idea here for someone who isn't deep into gravitational wave physics?
Tom: The core concept is using gravitational wave diffraction to map out how massive things are inside astrophysical objects, like dark matter halos or primordial black holes that we can't see directly. It connects the way the waves change frequency to the spatial structure of whatever is causing that distortion.
Lu: It’s essentially setting up a mathematical link where frequency evolution samples different spatial scales of the lens density, which is what makes it useful for probing internal structure rather than just measuring total mass.
Jane: That sounds complex. How do they actually make this connection mathematically? What’s the starting point for this entire framework?
Tom: They start with a linear order approximation in the lens potential, which directly relates the central diffractive image contribution to the Fourier transform of that projected lens density.
Meng: But then you have to deal with finite mass lenses, and that brings in this thing called a normalized mass form factor, which is what captures structure beyond just the total integrated mass.
Lalam: That form factor is key because it’s analogous to the ones you see in nuclear scattering experiments, allowing them to use those established concepts to model gravity effects.
Jane: Okay, so they connect wave frequency evolution across spatial scales to this form factor idea. What happens when you look at how these signals behave over time?
Tom: They focus on chirping gravitational waves, where the frequency changes over time, because that change is what lets the signal sample structure across different spatial scales.
Lu: Specifically, they use a relationship like k star equals x s divided by x two F to show how the frequency evolution samples the mass form factor at different spatial scales.
Meng: When you move from just looking at frequency to looking at time, that chirping signal is what enhances the ability of the gravitational wave to imprint that lens structure onto itself.
Jane: That’s interesting. So, what are they actually trying to distinguish? If we have an extended dark matter object versus a simple point mass, how does this method separate them?
Tom: They find that extended dark matter objects are most easily distinguished from point masses when their characteristic radii and scales line up near the Fresnel scale, specifically when x F is about x E about R.
Lu: For next-generation detectors like Cosmic Explorer or Einstein Telescope, they suggest this distinction becomes achievable for extended dark matter lenses with characteristic masses near ten squared M and radii around zero point one pc.
Meng: And for LISA and Taiji, the reach extends to characteristic masses of ten to the power of seven M and radii of about ten pc, which is a pretty big leap in scale.
Lalam: The strongest structural difference appears near the critical curve radius when it matches the peak frequency scale, which is where they get their best structural signal.
Jane: That gives us a real observational target. What about primordial black holes? Is this method unique for those objects?
Tom: They use this framework to probe primordial black holes especially when they are surrounded by dark matter halos, because matching the geometric optics observables to a bare black hole isolates the contribution from that surrounding halo structure.
Lu: This lets them look at mixed scenarios involving both macroscopic primordial black holes and microscopic particle dark matter because the halo structure is concentrated toward lower frequencies.
Meng: The detection limits they project for primordial black holes, under detection-only criteria for Cosmic Explorer and Einstein Telescope configurations, suggest a five-year limit around ten to the power of six for PBH about f.
Jane: That’s a very concrete number. And what about probing something even smaller, like ultralight bosons? Can they find those signatures?
Lalam: The framework is also extended to probe ultralight particle physics through the signature produced by scalar superradiance clouds around black holes, which can be distinguishable from matched bare black holes.
Tom: For those scalar states, the corresponding boson mass ranges they're looking at are approximately ten to seventeen to ten to the power of two eV.
Lu: Vector and tensor states show slightly different ranges, around ten to sixteen and ten to the power of-four eV respectively, which gives them a way to characterize ultralight particle physics.
Jane: So, if we put it all together for the Gravitational Wave Diffractive Images Reveal Structure of Astrophysical Objects paper, what is the big picture change this has for us listeners who don't read these papers?
Tom: The big picture is that gravitational wave diffraction isn't just about measuring how much mass there is; it’s a new way to look inside objects by using frequency changes to see their internal density distribution.
Meng: It moves us from just seeing the total gravitational effect to actually resolving the shape of the halo or the structure of a black hole.
Lalam: I am confident in this direction, as it provides a novel gravitational probe connecting wave distortions directly to spatial Fourier components of projected density.
More episodes
- 2603.28113-The multiplicative Lipschitz bound as an inductive bias
- 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