AI & Machine Learning

Provides a geometric 'manifold envelopment' framework to explain why unsupervised RL for mathematical reasoning often collapses and how to stabilize it.

Formalizes AI agent governance as 'policies on paths,' moving from static prompts to runtime enforcement of complex legal and safety constraints.

Enables stable 4-bit microscaling (MXFP4) quantization for Multi-modal LLMs, which previously suffered from performance collapse.

Introduces a way to train Reward Models that generate 'transferable rubrics'—explicit scoring criteria that improve performance across different tasks and models.

OmniSONAR scales cross-lingual sentence embeddings to over 1,500 languages across text, speech, code, and math in a single semantic space.

Aligns a base model to a target model's behavior by optimizing the 'data mixture' weights instead of using RLHF or DPO.

Achieves high-bandwidth, precise Cartesian control of a fully soft continuum robot, breaking the assumption that softness and precision are incompatible.

Fine-tuning language models on journal publication records allows them to match or exceed human experts in judging 'scientific taste'—the ability to identify which research ideas are worth pursuing.

This paper introduces a Markov-based discrete reasoning model that learns its own stopping criterion and can re-mask and correct its own mistakes.

Fast-WAM proves that World Action Models do not actually need to generate future 'imagination' frames at test-time to achieve state-of-the-art performance in embodied control.

The study provides a formal link showing that internal 'world model' representations in transformers are a direct byproduct of the predictive geometry of the training data.

Chain-of-thought (CoT) reasoning in Vision-Language Models systematically degrades the reliability of uncertainty estimates, making models dangerously overconfident.

Low-precision optimizer states cause 'state staleness' where updates round back to stored values, but scheduled resets can fully recover performance loss.

IQuest-Coder-V1 introduces a series of high-performance code models including a unique 'Loop' variant with a recurrent mechanism for efficiency.

This method non-rigidly aligns inconsistent video diffusion frames into globally-consistent 3D pointclouds to enable high-quality environment reconstruction.

Infrastructure-taught 3D perception uses static roadside sensors as unsupervised teachers for moving vehicles, eliminating the need for manual labels.

pADAM is a unified generative framework that learns shared priors across heterogeneous multi-physics families (e.g., scalar diffusion to Navier-Stokes).

The SOMP attack demonstrates that private training text can be reconstructed from shared gradients even at high batch sizes (up to B=128).

TraceR1 uses a two-stage reinforcement learning framework to train multimodal agents to forecast entire trajectories before execution, rather than acting reactively.

Zero-shot sim-to-real transfer for complex robotic manipulation is achievable using only synthetic simulated data at scale.

Video models perform reasoning during the diffusion denoising steps rather than sequentially across video frames.

Using the best-performing models as anchors for 'LLM-as-a-judge' evaluations significantly reduces the reliability of human ranking correlations.

GIST achieves O(N) complexity for Graph Transformers while maintaining gauge invariance, enabling scaling to meshes with 750K nodes.

Intermittently resetting an agent to a fixed state significantly accelerates policy convergence in Reinforcement Learning.

SOMA provides a unified, differentiable layer that bridges incompatible human body models like SMPL and SMPL-X in a single closed-form pass.

LEAFE allows LLM agents to internalize feedback as actionable experience, enabling them to backtrack and recover from failures autonomously.

Pretrained 3D generative models can be repurposed for high-quality part segmentation using less than 1% of the typical labeled data.

Neural PDE solvers are not learning general operators, but rather a family of solutions specifically indexed to the boundary conditions seen during training.

DreamPlan fine-tunes Vision-Language planners entirely within the 'imagination' of a video world model, bypassing costly physical robot trials.

SurgΣ is a massive open-source release of 5.98M multimodal conversations and foundation models for surgical intelligence.

Turns out the math for how things cool down or rot works just fine even if time doesn't move forward.

Our computers are way slower than they should be because they're hardwired to think time only goes one way.

Your satellite internet doesn't actually care about clouds—it’s just the hidden liquid water inside them that’s killing your signal.

If someone hacks a self-driving car, the way it steers leaves a 'fingerprint' that's so weird the car can actually tell it's being hijacked.

An AI just started cracking math problems about the laws of physics that have basically been bullying scientists for centuries.

There’s this 'impossible' crystal structure that lets you squeeze data down as small as you want without it ever breaking.

There's this one weird number—the natural log of 3—that basically decides if a group will work together or descend into total chaos.

When vanilla prices skyrocketed, farmers in Madagascar actually cleared *more* forest, killing the idea that getting richer helps the environment.

The main tool we use to decide if science is 'true' was actually just a lazy shortcut invented to deal with all the new scientists after WWII.

Diffusion LLMs can match autoregressive (AR) reasoning performance by using AR-generated plans as globally visible scaffolds.

Researchers identified just three specific attention heads that govern persona and style, enabling precise steering without degrading model coherence.

Factual selection in LLMs is driven by rotational dynamics on a hypersphere rather than scalar magnitude shifts, with the behavior emerging suddenly at the 1.6B parameter mark.

The Spherical Kernel Operator (SKO) replaces dot-product attention with ultraspherical polynomials to bypass the saturation phenomenon that bottlenecks world models.

Truncated-Reasoning Self-Distillation (TRSD) allows models to maintain accuracy even when their chain-of-thought traces are heavily shortened.

Sparse Autoencoders (SAEs) can be used to build retrieval models that outperform traditional vocabulary-based sparse retrieval in multilingual settings.

The ICaRus architecture allows multiple different models to share a single, frozen KV cache for the same prompt.

Using parallel associative scans achieves a 44x speedup in training continuous-time Spiking Neural Networks (SNNs).

RelayCaching eliminates redundant prefill computation in multi-agent systems by reusing the decoding-phase KV cache from previous agents.

ICPRL enables vision-language models to acquire physical intuition and adapt their policies in-context through trial-and-error interaction.

Prism prevents 'diversity collapse' in self-evolving reasoning systems by using semantic partitioning to guide the generation of new problems.