ThinkLLM
ModelsCapabilitiesUse CasesBenchmarksPapersGlossary
ModelsCapabilitiesUse CasesBenchmarksPapersGlossary
AboutPrivacyTermsRSS

ThinkLLM

Spot an error in our data? Let us know.

Papers

Recent AI research papers with accessible summaries. Updated daily from arXiv, summarized for developers who don't read papers regularly.

2455 papers12 this month12 topics
AllEvaluation 43Agents 42Applications 39Efficiency 29Reasoning 29Training 27Safety 23Multimodal 18Architecture 9Alignment 8Data 4scaling 2

Nov 16 – Nov 22(3)

TideCache: Efficient Hybrid LLM Serving with Transparent Cache Management and Layer-Wise Streaming

Nov 19, 2026

Jinbin Luo, Xiaozhe Yao, Shang Yang et al.

By treating different cache types as flexible pools and overlapping data transfers with computation, TideCache achieves 2x more concurrent requests and 20-59% lower latency—showing that intelligent memory management, not just bigger GPUs, unlocks efficiency gains.

TideCache is a serving system that optimizes GPU memory usage for hybrid LLM architectures by intelligently managing different cache types across GPU and CPU memory. It uses transparent page redistribution, smart prefetching during computation gaps, and adaptive request scheduling to fit more requests and reduce latency on CPU-GPU systems.

efficiencyarchitecturetraining

ProtoDiffusion: Token-per-Layer Generative Modeling of Multi-Layer Images

Nov 19, 2026

Adrià Ciurana Lanau, Harleen Hanspal, Lusine Abrahamyan et al.

By sharing a spatial prototype across layers and using one compact vector per layer, ProtoDiffusion achieves efficient multi-layer image generation that's 19× faster than decomposing flat images, while maintaining high reconstruction quality.

ProtoDiffusion generates multi-layer graphic designs (like Photoshop files) by representing each layer compactly as a single vector, rather than full spatial grids. This reduces memory from N×H×W tokens to H×W+N tokens. The model can generate designs from images or text, automatically determining layer count and layout without requiring them as input.

Nov 9 – Nov 15(6)

Unbundling Multimodal Vision Encoding

Nov 12, 2026

Ellis Brown, Chris Hoang, Shengyi Qian et al.

For building multimodal models, dedicated vision capacity trained jointly with the language model is more important than using pretrained vision encoders, and this capacity can be efficiently placed early in the backbone rather than as a separate pre-fusion module.

This paper investigates how vision encoders should be designed in multimodal language models by systematically comparing three design choices: using a pretrained visual encoder, dedicating capacity to vision processing, and where to place that capacity.

multimodalarchitecturetraining

ArchSpace: Recording, Reviewing, and Referencing LLM Architecture Research

Nov 12, 2026

Jie Wang, Tong Zhu, JiaHao Kuang et al.

ArchSpace enables researchers to test LLM architecture ideas efficiently by predicting large-scale performance from small experiments using scaling laws, creating a shared record of what works and why.

ArchSpace is a public platform for documenting, evaluating, and sharing LLM architecture innovations. It addresses a gap where industry teams test many architectural ideas but rarely publish which succeeded or failed, and academia struggles with small-scale comparisons.

Nov 2 – Nov 8(3)

BitDecoding-FP4: Native FP4 KV Cache Decoding via Algorithm-System Co-Design

Nov 6, 2026

Dayou Du, Woon Yee Ng, Guangming Sheng et al.

You can compress LLM key-value caches to 4-bit precision without significant accuracy loss by carefully handling quantization errors at the algorithm level and co-designing with hardware capabilities—delivering major serving speedups in practice.

BitDecoding-FP4 enables efficient long-context LLM serving by storing key-value caches in ultra-low-precision 4-bit floating-point format while maintaining accuracy.

efficiencytrainingarchitecture

Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout

Nov 5, 2026

Tushaar Gangavarapu, Gil Tabak, Clemens JS Schaefer

Linear attention models can be made significantly more efficient by allocating parameters based on sensitivity analysis: use low-rank gates for state erasure (which matters less) and invest savings into wider networks and better readout mechanisms.

Muir improves linear attention models by analyzing how state updates work differently across channels. The key insight is that erasing old information requires less capacity than writing new information, so the model uses low-rank gates for erasure and reinvests those parameters into a wider network.

Oct 26 – Nov 1(12)

Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention

Oct 30, 2026

Ales Kubicek, Andrea Caciolai, Liangzu Peng et al.

LSA enables practical ultra-long context (1M+ tokens) by replacing expensive full attention with hierarchical cluster scoring, achieving 22.7x speedup in prefill and 4.6x in decode while preserving model accuracy.

This paper introduces Latent Sparse Attention (LSA), a method that makes attention efficient at ultra-long context lengths (1M+ tokens) by hierarchically clustering keys in a learned latent space. Instead of comparing queries to every key, queries score clusters of keys, reducing computation from quadratic to subquadratic while maintaining accuracy.

efficiencyarchitecturescaling

ACCT: Asymmetric Cross-LLM Context Transfer

Oct 30, 2026

Yuzhe Liu, Haolin Tian, Tonghan Wang

You can transfer pre-computed context between different-sized models in the same family, cutting context processing time by 6-21x—useful when deploying multiple model sizes and reusing the same input across them.

ACCT enables reusing computed context (KV cache) across different LLM sizes within the same family, avoiding redundant computation when switching models. It transforms a source model's KV cache into sparse, compressed states that a smaller target model can directly use, achieving 6.5-21x speedup on context processing while maintaining performance.

Oct 19 – Oct 25(10)

Beyond Discrete Tokens: Preserving Numerical Continuity in Medical Vision-Language Models

Oct 23, 2026

Dianye Huang, Jiamei Guo, Yue Zhou et al.

Combining diffusion-based continuous prediction with language models outperforms discrete token prediction for medical measurements, suggesting hybrid architectures better match the continuous nature of real-world quantitative tasks.

This paper presents HAD-VLM, a medical AI system that combines language understanding with precise numerical measurement from ultrasound and X-ray images.

multimodalarchitecture

Low-Bit N-Gram Embeddings for Language Models

Oct 22, 2026

Stephen Chung, Wenyu Du

You can compress n-gram embedding tables to 4-bit precision and save 75% of their memory with minimal performance loss—or use the freed memory to make the model better overall.

This paper shows that n-gram embedding tables in language models can be stored at low precision (4 bits) instead of full precision, reducing memory usage by ~75% while keeping 99% of the performance benefit. The authors test quantization methods on models from scratch and on large existing models like DeepSeek, finding that saved memory can be reinvested to improve overall performance.

efficiency

Oct 12 – Oct 18(11)

Rethinking the design space of Vanilla Pixel Diffusion Transformers

Oct 16, 2026

Soumik Mukhopadhyay, Long Mai, Abhinav Shrivastava et al.

Pixel diffusion models benefit from adaptive prediction strategies that vary by capacity and data type, and post-hoc high-frequency refinement can significantly improve image quality without expensive retraining.

This paper improves pixel-space diffusion transformers by introducing MFlow, which adaptively learns to blend two prediction strategies (clean image vs. velocity) based on data characteristics and model size. The authors also propose ConvRefiner, a lightweight module that fixes a persistent high-frequency quality deficit in generated images without retraining the main model.

architectureefficiencytraining

Learning without relearning physical symmetry: an SO(3)-equivariant convolutional network (CNN) on the sphere

Oct 16, 2026

Joshua Williamson, Niall Jeffrey, Lorne Whiteway

Encoding physical symmetries directly into neural network architecture—rather than letting the model learn them—dramatically improves data efficiency and accuracy for scientific problems with inherent geometric structure.

This paper presents a spherical convolutional neural network designed for data naturally living on spheres (like cosmic microwave background maps). By building in rotational symmetry mathematically rather than learning it, the network uses training data more efficiently and achieves better results on cosmology tasks with fewer simulations needed.

Oct 5 – Oct 11(12)

On the Design Space of Continuous Diffusion Language Models

Oct 9, 2026

Oscar Davis, Daan Roos, Alice Bizeul et al.

Self-conditioning and careful embedding geometry are critical design choices for continuous diffusion language models; simple architectural improvements like DoubleCat can match or exceed discrete diffusion performance without additional computational cost.

This paper investigates how to design better continuous diffusion language models by studying two key factors: self-conditioning (exposing models to realistic inference states during training) and embedding geometry (how word representations are shaped).

trainingarchitectureefficiency

Bi-FORK: Generative Modeling of High-Dimensional Bifurcating Systems

Oct 8, 2026

Anna Zimmel, Fleur Hendriks, Markus Holzleitner et al.

Bifurcating systems violate the one-to-one assumption in standard neural surrogates; Bi-FORK solves this by treating bifurcations as a generative modeling problem, enabling amortized prediction of all solution branches simultaneously.

Bi-FORK is a generative model that learns one-to-many solution maps in high-dimensional physical systems undergoing bifurcations—where a single input produces multiple equally valid outputs. Using latent flow matching and repulsion-guided sampling, it generates complete trajectories while preserving spatial and temporal coherence, scaling to systems with hundreds of thousands of dimensions.

Sep 28 – Oct 4(17)

Embedding Prediction Helps Image Generation

Oct 1, 2026

Sihan Xu, Ji Xie, Zilin Wang et al.

Dynamically predicting and updating conditioning embeddings at each generation step improves diffusion model efficiency and quality—you don't need to reuse the same embedding throughout the entire denoising process.

This paper proposes using predicted embeddings as dynamic conditioning signals in diffusion transformers instead of static embeddings. A separate transformer (NEPA) predicts image embeddings at each denoising step, allowing the conditioning to adapt to the current noise level. The approach achieves competitive image generation quality on ImageNet with significantly less training compute.

architectureefficiencytraining

SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation

Oct 1, 2026

Tianjiao Yu, Xinzhuo Li, Yifan Shen et al.

Representing 3D shapes as overlapping 2D slices rather than voxels dramatically reduces computational cost while improving topological correctness—a practical win for efficient 3D generation at scale.

SILSA is a 3D generation framework that uses compact sliding-window slice representations instead of expensive voxel tokens to generate high-resolution 3D shapes. By encoding cross-sections along three axes and adding topology supervision based on persistence diagrams, it achieves better structural quality while using 70% fewer tokens and running 58% faster than competing methods.

Sep 21 – Sep 27(5)

EAServe: Encode-Aware Disaggregated Serving for Multimodal Large Language Models

Sep 25, 2026

Kunxiong Zhu, Zhihao Shu, Hangyu Zheng et al.

Multimodal LLM serving requires rethinking GPU resource allocation around the Encode stage—treating it as a bottleneck control point rather than a separate service unlocks significant throughput gains.

EAServe optimizes serving multimodal LLMs by treating the Encode stage as a control point for the three-stage Encode-Prefill-Decode pipeline. It uses adaptive micro-batching, partial offloading, and GPU partitioning to balance resource utilization across stages, achieving 4.3x higher throughput than existing systems under latency constraints.

efficiencymultimodalarchitecture

SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance

Sep 24, 2026

Xinyue Zeng, Jiawei Zhang, Yujun Yan et al.

Long-horizon reasoning failures in LLMs stem from structural biases in the reasoning space itself, not just model capacity—and injecting geometric structure into the reasoning process can dramatically improve performance on hard problems.

This paper addresses why large language models struggle with long-horizon reasoning tasks by identifying two key problems: exploration bias (getting stuck in locally plausible but structurally weak paths) and compounding bias (small errors accumulating over many steps).

Sep 14 – Sep 20(9)

Abstention and Noise Filtering: Two Missing Primitives of Softmax Attention

Sep 18, 2026

Richard Zhe Wang

Attention heads need both the ability to abstain from attending and to filter noise from values—their importance shifts with model scale, suggesting future architectures should support both primitives.

This paper identifies two missing capabilities in standard softmax attention: abstention (allowing heads to output nothing instead of always producing weighted combinations) and noise filtering (suppressing interference from mixed features).

architectureefficiencytraining

JEPA-Anything: Learning Predictive Models across Different Worlds

Sep 17, 2026

Taoyong Cui, Zhongyao Wang, Xinyue Xu et al.

A single learning principle based on factorized predictions can work across radically different domains, suggesting world modeling doesn't need domain-specific architectures—and can even guide real scientific discovery.

JEPA-Anything is a unified framework for building predictive models across completely different domains—from videos to molecules to weather—using a technique called orthogonal predictive factorization. Instead of training separate models for each domain, it learns to decompose predictions into independent factors that work across vision, biology, physics, and clinical data.

Sep 7 – Sep 13(4)

Distance generalization in transformers: why bother with positional encoding?

Sep 10, 2026

Daniel Henrik Nevermann, Claudius Gros

Positional encodings like RoPE and ALiBi don't automatically help transformers generalize to unseen token distances—data diversity and task structure matter more than the encoding scheme itself.

This paper investigates how transformers generalize to different token distances between training and inference, using synthetic copy tasks. It compares positional encoding schemes (RoPE, ALiBi, no encoding) and finds that understanding distance generalization requires rethinking how we use positional information.

evaluationarchitecture

From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge

Sep 10, 2026

Wenkang Wei, Yuan Fang, Renhe Jiang et al.

Language models have a critical handoff point where they transition from using query routing information to relying on internal knowledge—this happens at different layers across models and reveals how they internally organize and access information.

This paper investigates how large language models retrieve and use internal knowledge when answering questions by analyzing how different layers process query information versus stored knowledge.

Aug 31 – Sep 6(8)

UniMate: One Unified Model to Animate Diverse Skeletons

Sep 4, 2026

Linzhan Mou, Jiahui Lei, Zhiyang Dou et al.

For the first time, a single model can animate any skeleton topology (bipedal, quadrupedal, insects, etc.) from text alone—no per-character fine-tuning or reference motions needed at inference time.

UniMate is a foundation model that generates realistic motion for any 3D character skeleton from text descriptions, without needing to retrain for each new character type. It uses a specialized neural architecture that understands skeleton structure through graph-based attention mechanisms, and was trained on a diverse dataset of 13,000+ motion sequences across different creature types.

multimodalarchitecture

Reflection-aware Generative Novel View Synthesis

Sep 4, 2026

GeonU Kim, Shin Dong-Yeon, Tae-Hyun Oh

You can generate realistic novel views of mirror scenes by treating reflections as virtual views and using gated attention mechanisms—no retraining needed, just clever use of existing diffusion models.

This paper presents Ref-GeNVS, a method for generating novel views of scenes containing mirrors without requiring additional training. The key innovation is treating mirror reflections as complementary views by estimating the mirror plane and reflecting camera poses, then using a two-stage approach with special attention mechanisms to ensure reflections stay consistent during generation.

multimodalarchitectureefficiency

Atomic MathReasoner: Instance-Level Composition of Atomic Reasoning Skills for Mathematical LLMs

Nov 18, 2026

Jiashu Li, Tianyu Gao, Qinglong Liu et al.

Breaking math reasoning into specialized expert modules and routing problems to appropriate skill combinations outperforms monolithic approaches—AMR improves math accuracy by 3-5% on benchmark tasks by matching problems to their required reasoning skills.

This paper proposes Atomic MathReasoner (AMR), a framework that decomposes mathematical reasoning into specialized skills and learns to compose them for different problems. Rather than training one general math model, AMR creates expert modules for specific reasoning abilities, then uses a lightweight router to select the right combination of experts for each problem at inference time.

trainingarchitecturereasoning
architecturescalingevaluation

Classical Energies Learn to Reason 1,000x Faster than Looped Models

Nov 12, 2026

Benjamin Hoover, Hugh Blayney, James C. R. Whittington

Energy-based models offer a faster, more efficient alternative to looped reasoning architectures for constraint satisfaction problems, suggesting that iterative reasoning may be overkill for many tasks.

This paper shows that energy-based models (specifically a learned version of Hopfield networks) solve reasoning puzzles like Sudoku and Mazes 1,000x faster than looped reasoning models, while using far fewer computations. The approach learns to encode puzzle constraints directly into an energy function, enabling efficient problem-solving without iterative loops.

reasoningefficiencyarchitecture

Goal Encoding in LLMs

Nov 12, 2026

Arnab Sen Sharma, Eric Todd, Sheridan Feucht et al.

LLMs encode goals in sparse attention heads and pursue them by repeatedly re-reading task specifications rather than maintaining internal state, suggesting a fundamentally different mechanism than human goal-directed behavior.

This paper investigates how large language models internally represent and pursue goals. Researchers found that specific attention heads encode the next sub-goal, and this encoding causally influences the model's actions.

reasoningarchitecture

Symmetry in Time: A Representation Theory of Linear Recurrent Memory

Nov 12, 2026

Anif N. Shikder, Ramit Dey, Jan Minac et al.

SSMs can only track commutative groups when using diagonal real/complex updates; adding quaternion channels enables tracking of non-commutative groups like S₅ while keeping linear-time inference, and the learned weights become directly interpretable.

This paper uses representation theory to explain why standard state space models (SSMs) fail at tracking non-commutative sequences, and proposes HyperSSM, which adds quaternion-valued channels to handle complex group structures while maintaining efficient inference. The theory directly explains what SSM weights must encode, enabling interpretable learning of geometric transformations.

architecturereasoning

Asymmetric Multimodal Learning with Decomposed JEPAs

Nov 12, 2026

Yusuf Shihata

For multimodal AI systems, alignment-only training catastrophically fails on compositional reasoning when modalities carry different types of information—you need to explicitly preserve modality-specific structure alongside alignment.

This paper shows that standard cross-modal alignment alone loses important information when modalities are asymmetric (like text queries and images). It introduces Decomposed JEPA, which splits learned representations into a shared aligned part and private modality-specific parts, recovering information that alignment discards and improving performance on reasoning tasks.

multimodalarchitecturetraining
efficiencyarchitecturetraining

CATF: A Manager-Worker Framework for Context-Aware Time-Series Forecasting

Nov 3, 2026

Yueqiao Chen, Dongyu Liu

Hard routing through a Manager model lets specialized Workers handle different data regimes automatically, improving forecasting accuracy while reducing inference costs compared to soft-routing ensemble approaches.

CATF is a forecasting framework that uses a Manager-Worker architecture to predict time-series values when multiple contextual factors (like weather or calendar) influence the data. A Transformer-based Manager intelligently routes each input to a specialized Worker model based on context, allowing workers to automatically specialize without manual data splitting.

architectureefficiencytraining
efficiencyarchitecture

Subquadratic Attention with a Hierarchical Tree-Structured Indexer

Oct 30, 2026

Leon Lufkin, Cutter Dawes, Alexander Speicher et al.

ZHA achieves subquadratic attention scaling by using a hierarchical tree indexer, potentially enabling LLMs to handle contexts of hundreds of millions to billions of tokens—a major step toward solving the long-context bottleneck without sacrificing expressiveness.

This paper introduces Zyphra Hierarchical Attention (ZHA), a method that reduces the computational cost of attention in large language models from quadratic to subquadratic scaling.

efficiencyarchitecturescaling

Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers

Oct 29, 2026

Jiongfan Yang, Weichen Fan, Haiwen Diao et al.

By splitting image frequency components across local and global attention paths, you can generate finer image details without the quadratic cost of full global attention—achieving better quality at lower computational cost.

SpecSlot is a pixel-space image generation model that separates low-frequency and high-frequency image details into different processing paths. Low-frequency information uses global attention for context, while high-frequency details are processed locally, reducing computational cost while maintaining image quality.

architectureefficiencymultimodal

An ODE-based Sampler for Discrete Diffusion Models

Oct 29, 2026

Samson Gourevitch, Yazid Janati, Eric Moulines et al.

Discrete diffusion models can now use deterministic ODE-based sampling (like continuous diffusion models do), enabling faster inference and better distillation without sacrificing generation quality.

This paper develops a deterministic sampling method for discrete diffusion models by introducing exponential residual flows (ERFs), which use ODEs to guide token generation.

efficiencytrainingarchitecture

Multi-View MoE: Letting Every Router Shape Expert Computation

Oct 29, 2026

Zelin Zhao, Tianchi Ma, Junyuan Shang et al.

Using multiple routers with different expert preferences improves expert selection in MoE models, and reusing routing signals to customize expert inputs boosts performance without increasing compute cost.

This paper improves sparse mixture-of-experts models by using multiple routers to select experts and modulate their computation. Instead of one router deciding which experts to use, multiple independently-trained routers vote on expert selection, and their preferences are reused to customize each expert's input representation.

architectureefficiencytraining

NEOPIC: A Neural Operator Framework for Particle-based Kinetic Plasma Simulations

Oct 29, 2026

Sriramkrishnan Muralikrishnan, Chelsea Maria John, Jeffrey Kelling et al.

Neural operators can efficiently replace expensive field solvers in particle simulations by learning directly from particle data, with proven error bounds and linear computational complexity independent of mesh resolution.

NEOPIC is a neural operator framework that learns to replace the field solver in particle-based plasma simulations. Unlike grid-based approaches, it operates directly on particle positions without requiring a mesh, avoiding noise aliasing issues. The framework is plug-and-play, works with different simulation codes, and comes with theoretical guarantees on accuracy and computational complexity.

architectureefficiencyreasoning

Heads of Security: Learned Attention Gates Improve Prompt Injection Robustness

Oct 27, 2026

Igor Shilov, Xiaoxue Yang, Alexander Panfilov et al.

By learning which attention heads should focus on instructions vs. user input, you can defend against prompt injection attacks with minimal overhead—HoS reduces attack success rates from 97% to 10% while preserving the model's ability to use tools and answer questions correctly.

Heads of Security (HoS) is a defense against prompt injection attacks that works by training small attention gates for each transformer head to separate trusted instructions from untrusted user data. Unlike methods that require full model retraining, HoS adds minimal parameters while maintaining model utility and resisting both black-box and white-box attacks far better than existing defenses.

safetyarchitecturetraining

Attention as Lattice Gauge Theory: Relations Interact Only Through Wilson Loops

Oct 27, 2026

Aakash Sarkar, Ananya Mukherjee, Motoshige Sato

Attention mechanisms aren't arbitrary design choices—they're constrained by symmetry principles from physics.

This paper reveals that attention mechanisms in neural networks follow the mathematical structure of gauge theory when entities can rotate independently. The authors show that relations between entities can only interact through closed loops (Wilson loops), leading to a unique form of attention called holonomy attention.

architecturereasoning

Multi-State Linear Attention with Generalized Second-Order Routing

Oct 27, 2026

Hee Seung Hwang, Sanghyuk Chun, Xindi Wu et al.

By incorporating variance information into how segments are aggregated in linear attention models, MERLIN enables better long-context understanding without the quadratic complexity of standard attention—useful for scaling LLMs to longer sequences.

This paper improves Multi-State Linear Attention (MSLA) for long-context language models by tracking both the mean and variance of key alignments across segments, rather than just the mean.

efficiencyarchitecturereasoning

GAS-PFN: Graph-Aware Spatiotemporal Prior-Data Fitted Networks for Zero-Shot Traffic Forecasting

Oct 27, 2026

Joseph Natter, Long Vu, Yifan Zhang et al.

By explicitly encoding road network structure into the training process and using a two-stage inference design, you can build traffic forecasters that work on new networks without retraining—a major practical advantage for deploying to evolving city infrastructure.

This paper introduces GAS-PFN, a neural network that forecasts traffic flow without needing to be retrained for new road networks. It learns from synthetic traffic patterns during training, using graph structure (road topology) to understand how traffic moves between connected sensors.

reasoningarchitectureapplications

Work Agents: Systems, Capabilities, and Optimization

Oct 26, 2026

Zishan Xu, Enjun Du, Mingqi Gao et al.

Work agents require a system-level perspective combining foundation models with runtime infrastructure, and their reliability depends on five core capability families and joint optimization of both the model and execution harness.

This paper presents a comprehensive framework for understanding work agents—AI systems that execute complex tasks across tools and applications.

agentsarchitectureevaluation
training
architecture

Rethinking Sequence Compression with Test-time Regression

Oct 22, 2026

Sihan Yang, Jie Huang, Hanxuan Li et al.

Test-time regression offers a more effective way to compress token sequences than standard pooling or projection methods, enabling better efficiency in both language and multimodal models without sacrificing performance.

This paper proposes TTR-Compressor, a new method for compressing sequences of tokens in language and multimodal models. Instead of using pooling or learned projections, it frames compression as a regression problem solved at test time with coefficients computed efficiently in closed form.

efficiencyarchitecturemultimodal

UniWALM: Unified World Action Language Model for Embodied Robot Control

Oct 20, 2026

Rahul Sajnani, Diego Romeres, Mingyuan Zhang et al.

Integrating language generation directly into robot control models—not as an afterthought but as a core part of the same network—enables robots to use language as working memory, significantly improving performance on tasks that require remembering past events.

UniWALM is a unified AI model for robot control that learns video, actions, and language together in one shared representation.

agentsmultimodalarchitecture

Efficient Edge-aware Attention Network for Graph Generation

Oct 20, 2026

Vincent Jung, Alba Carballo-Castro, Yiming Qin et al.

Graph generative models can achieve significant memory and speed improvements by redesigning attention to work with hardware-optimized implementations, making graph foundation models more practical to scale.

This paper proposes an efficient attention mechanism for graph generation models that reduces GPU memory usage by 65% while maintaining generation quality. The key innovation is a head-wise score modulation approach that incorporates graph structure into attention, implemented using PyTorch's FlexAttention to leverage modern hardware optimizations.

efficiencyarchitecturescaling

Turning TS2Vec into a Foundation Model for Time Series Classification

Oct 20, 2026

Vasilii Feofanov, Zhenghan Tai, Andrei Ermakov et al.

Time series foundation models can match specialized methods without attention mechanisms or complex tokenization—a simple convolutional encoder with dual-view inputs and synthetic pre-training is sufficient.

Researchers adapted TS2Vec, a self-supervised time series method, into a reusable foundation model by pre-training on synthetic data, using dual input views (raw amplitude and normalized shape), and distilling two specialist encoders into one.

architecture

Evolving Agent Interfaces

Oct 20, 2026

Saaket Agashe, Yuzhe Yang, Gonzalo Gonzalez-Pumariega et al.

Interface design can be automated and improved through agent experience—evolving interfaces from task trajectories is a practical way to boost agent performance that's complementary to model scaling.

This paper introduces Agent S4, a system that automatically designs and improves computer interfaces for AI agents by learning from task experience. Rather than relying on hand-built interfaces, a meta-agent analyzes how task agents perform on training tasks and evolves better interfaces and usage strategies.

agentsarchitecturetraining

LatComm: Reusable Context Compression for Latent Communication between Large Language Models

Oct 19, 2026

Huawei Lin, Jinting Liu, Tan Yu et al.

Separating context encoding from reading allows LLM systems to compress context once and reuse it across multiple questions with different models, dramatically reducing computation and latency in multi-turn scenarios.

LatComm enables efficient communication between multiple LLMs by compressing context once into a compact, reusable representation that different models can selectively read.

efficiencymultimodalarchitecture

Kronecker Delta Rule for Long Context Linear Attention

Oct 19, 2026

Anamaria-Roberta Hartl, Levente Zólyomi, Michael List et al.

The Kronecker Delta Rule offers a theoretically grounded way to update associative memories in attention models that's both stable and efficient for long-context tasks, outperforming existing linear attention methods.

KronLSTM introduces a new memory update mechanism for linear attention models that modifies memories along a single direction, reducing interference with stored information. This approach provides mathematically stable memory dynamics without extra normalization, making it effective for processing long documents and sequences.

architectureefficiencyreasoning

Competing Learning in Hybrid Architectures

Oct 19, 2026

Xingyu Shen, Yingfa Chen, Ziqing Qiao et al.

In hybrid architectures mixing recurrence and attention, recurrence dominates learning of retrieval functions. Staged training that temporarily freezes recurrence allows attention to develop, significantly improving model performance.

This paper identifies a 'Competing Learning' problem in hybrid language models that combine recurrent and attention layers: recurrence learns retrieval functions faster, blocking attention from learning. The authors propose staged training—freezing recurrence updates mid-training—to let attention catch up, improving performance on language modeling and retrieval tasks.

trainingarchitectureefficiency
architectureefficiencyreasoning

Discovering state-of-the-art quantum circuits with large language models

Oct 15, 2026

Andi Gu, Alek H. Kemeny, Susanne F. Yelin et al.

LLMs can effectively co-design quantum algorithms by proposing circuit structures that an evaluator scores, discovering practical improvements over hand-optimized quantum code without requiring the LLM to understand quantum mechanics deeply.

Researchers used large language models working with an independent evaluator to discover optimized quantum circuits for various computational tasks.

reasoningarchitectureapplications

Rethinking Routing Exploration in Modern Mixture-of-Experts

Oct 15, 2026

Bingcong Li, Hao Ma, Michael Muehlebach et al.

When designing MoE models with many experts and shared components, you need to actively manage routing exploration—suppressing unnecessary experts per token can help the model find better expert combinations without getting lost in the combinatorial space.

This paper identifies a fundamental tension in modern Mixture-of-Experts models: fine-grained expert segmentation requires more exploration of the routing space, but shared experts reduce the variability needed for that exploration. The authors propose MirrorMoE, which uses token-adaptive expert suppression to balance these competing needs and improve performance across different model scales.

architecturescalingefficiency

GeoState: Learning View-Decoupled Geometry States for 3D World Modeling

Oct 15, 2026

Hongchang Liu, Jian Liu, Yujuan Ding et al.

By decoupling scene state from camera viewpoint, GeoState makes 3D world predictions more accurate and generalizable—the same predicted scene can be rendered from different camera angles without retraining.

GeoState is a 3D world modeling approach that separates scene geometry from camera viewpoint by learning view-decoupled state representations. It uses a scene tokenizer to aggregate multi-view geometry features into a unified state, then predicts future states with an autoregressive model while handling camera motion separately.

architecturereasoningmultimodal

Designing Transformers for Continuous Diffusion Language Models

Oct 15, 2026

Weijian Luo, Yujian Chen, Hao Jiang et al.

Transformer architecture design matters as much as training objectives for diffusion-based language models—using architectures optimized for CDLMs rather than borrowed from LLMs or image diffusion models substantially improves generation quality.

This paper redesigns Transformer architectures specifically for continuous diffusion language models (CDLMs), which generate text by gradually refining noise into coherent sequences.

architecturetrainingefficiency

ACEdit: Stable Multi-Turn Image Editing via Asymmetric Conditioning

Oct 15, 2026

Sara Dorfman, Nir Goren, Gal Barak et al.

Multi-turn image editing fails because each edit shifts the image distribution away from training data. Training with self-forcing and asymmetric teacher-student conditioning keeps quality stable across many sequential edits.

This paper tackles quality degradation in multi-turn image editing by training models to handle their own outputs from previous edits. The key innovation is an asymmetric conditioning scheme where a student model edits sequentially on its own outputs while a teacher model stays anchored to the original image, using distillation to provide reliable training signals despite distribution shifts.

trainingarchitectureefficiency

Twice Upon a Time: Complementary Representations for Time-Series Reasoning

Oct 15, 2026

Hedi Zisling, Liran Nochumsohn, Ilan Naiman et al.

Combining multiple complementary representations of the same time-series data—each with different invariance properties—helps language models better capture both absolute values and temporal structure, improving performance on prediction and reasoning tasks.

This paper proposes using two complementary visual representations for time-series reasoning in language models: line plots that preserve raw measurements, and normalized delay embeddings that expose temporal patterns. A dual-view model combining both through visual encoders significantly outperforms single-representation baselines on time-series understanding benchmarks.

multimodalreasoningarchitecture

Modeling and Understanding Neural Weights of Humanoid Policies with Diffusion

Oct 14, 2026

Huixuan Cheng, Bike Zhang, Shankar Sastry et al.

You can generate functional robot control policies by modeling the distribution of neural network weights themselves, and the learned weights naturally encode the robot's physical structure regardless of how they were trained.

This paper shows that neural networks controlling humanoid robots can be generatively modeled directly through their weights using diffusion models. The authors compress robot control network parameters into a latent space and train a diffusion transformer to model the distribution of these latents.

architecturetraining

Your Transformer Is Secretly Elastic

Oct 14, 2026

Walter Simoncini, Dheeraj Varghese, Gertjan J. Burghouts et al.

You can prune any pre-trained transformer to run at variable speeds without retraining by jointly ranking all neurons and heads on a common scale—enabling one model to adapt to different hardware and latency constraints.

This paper introduces a method to make pre-trained transformers 'elastic'—able to dynamically adjust their computational cost on-the-fly without retraining. By jointly ranking individual neurons and attention heads across the entire model and optimizing under soft pruning at multiple sparsity levels simultaneously, the approach enables smooth performance degradation from full to 60% sparsity.

efficiencytrainingarchitecture

Continuum Attention

Oct 14, 2026

Afshin Khadangi

Continuum Attention replaces discrete token lookup with continuous integration over memory, improving language modeling performance without increasing computational cost—suggesting continuous functions may be a better foundation for attention than discrete token operations.

This paper introduces Continuum Attention, a new attention mechanism for Transformers that treats memory as a continuous curve instead of discrete tokens. Rather than looking up individual key-value pairs, the mechanism integrates over a smooth memory representation weighted by exponential relevance.

architectureefficiencytraining
architecturetrainingapplications

One Block, Multiple Depths: Recurrent Vision Transformers with Depth-Programmed Experts

Oct 8, 2026

Adrian Bulat, Yassine Ouali, Georgios Tzimiropoulos

A single transformer block with depth-programmed expert mixtures can replace deep vision encoders, reducing parameters dramatically while maintaining accuracy—useful for efficient vision models and elastic deployment at multiple depths.

This paper proposes reViT, a vision transformer that uses a single block applied repeatedly instead of stacking many layers. By representing the feed-forward network as a mixture of shared experts controlled by depth coordinates, it matches full-depth encoders with 70% fewer parameters and comparable computation. The approach works both for training from scratch and distilling from teacher models.

efficiencyarchitecturetraining

Beyond Spatio-Temporal Priors: A Generalizable Approach for Dense Correspondence Matching

Oct 8, 2026

Luping Liu, Bingyi Kang, Yifan Wang et al.

By combining foundation models with diverse training data, you can build correspondence matching systems that work across both classical vision tasks and modern generative applications like image editing.

FreeMatching is a framework for matching visual correspondence across images that goes beyond traditional motion and geometry assumptions. It combines generative and semantic models trained on diverse data sources to handle challenging transformations in image editing and generation tasks, while maintaining competitive performance on standard benchmarks.

multimodalevaluationarchitecture

Learning Kilometer-Scale Weather Prediction with Global-Regional Alignment

Oct 8, 2026

Guowen Li, Yang Liu, Yujie Wang et al.

You can build accurate kilometer-scale weather forecasts by aligning global and regional model representations, allowing a single regional model to work with multiple pretrained global forecasters without retraining.

ScaleCast is a weather forecasting framework that combines global weather models with high-resolution regional predictions by aligning their representations across different grid scales. It enables kilometer-scale weather forecasts for local warnings while reusing pretrained global models, eliminating the need to retrain when switching between different global forecast sources.

architecturemultimodalscaling

SpaceFlow: Locally Controllable 3D Generation

Oct 8, 2026

Neil De La Fuente, Joan Lafuente, Mukhammadali Sayfiddinov et al.

Users can now control different regions of generated 3D objects independently—some parts can follow input geometry exactly while others generate freely—solving the problem of global-only control in current 3D generation methods.

SpaceFlow enables fine-grained control over 3D object generation by letting users specify which parts should strictly match input shapes versus allow creative variation. It combines geometric primitives as spatial constraints with text/image conditioning to control both structure and appearance locally, avoiding unwanted blending between object parts.

multimodalarchitectureapplications

HRIL: Learning Multimodal Synergy via Higher-Order Tensor Modeling

Oct 8, 2026

Qun Dai, Liangjian Wen, Jiang Duan et al.

Synergistic multimodal information requires explicit higher-order tensor modeling rather than standard pairwise interactions; HRIL's approach consistently outperforms existing methods on tasks where modalities must work together.

This paper addresses how to capture synergistic information in multimodal learning—signals that only emerge from combining multiple modalities together, not from any single modality alone. The authors propose HRIL, which uses higher-order tensor mathematics (specifically Tucker decomposition) to model complex cross-modal interactions and preserve the capacity to learn these synergistic signals.

multimodaltrainingarchitecture

Marformer: A Transformer for Predicting Missing Data Distributions

Oct 8, 2026

Prabhav Singh, Xiheng Tom Wang, Haojun Shi et al.

For incomplete data problems, directly predicting conditional marginals with a Transformer can be faster and more accurate than traditional statistical methods or full generative models, enabling better decision-making under uncertainty.

Marformer is a Transformer model that predicts conditional probability distributions over missing variables given observed data. Unlike generative models, it directly outputs marginal distributions needed for decision-making under uncertainty, without modeling the full joint distribution. It's trained like BERT to predict missing values and works in a single forward pass.

architecturetrainingevaluation

4D-HOF: Hand-Object Flow Matching for Feed-Forward 4D Interaction Reconstruction

Oct 6, 2026

Shiqi Li, Sean Cho, Yijie Li et al.

Feed-forward generative models can reconstruct complex hand-object interactions faster and more reliably than optimization-based methods by learning to correct foundation model errors while respecting physical constraints during generation.

This paper presents 4D-HOF, a fast method for reconstructing 3D hand and object positions/orientations from video. Instead of slow per-video optimization, it uses a generative model trained on diverse data to refine rough estimates from vision models.

architectureefficiencyreasoning

Denoising Hierarchical Representations: Joint Continuous Diffusion for Language Modeling

Oct 6, 2026

Mathias Ollu, Nikos Komodakis

Diffusing tokens at multiple granularities (fine-grained and clustered) in parallel improves continuous diffusion language models substantially, achieving state-of-the-art results on text generation and reasoning benchmarks.

This paper introduces Hierarchical Continuous Diffusion Language Models (H-CDLMs), which improve text generation by diffusing tokens at multiple semantic levels simultaneously—both individual tokens and coarser token clusters. Applied to existing models like CoBit and FLM, this approach significantly boosts generation quality and reasoning performance with minimal computational overhead.

architecturetrainingefficiency

Towards Looped Models Done Right, Part II: Rethinking at Fixed Points

Oct 5, 2026

Benhao Huang, Chufan Shi, Junlin Chen et al.

Looped models can be dramatically more efficient by recognizing that recurrent states converge to fixed points, enabling truncated backpropagation, KV cache sharing, and faster RL—with learned depth priors and orthogonal injection providing better supervision than existing approaches.

This paper optimizes looped language models—models that process information through multiple recurrent passes—by leveraging fixed points in recurrent states.

trainingefficiencyarchitecture

Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model

Oct 5, 2026

Sahil Mahendrakar

Knowledge distillation can compress speech synthesis models by 10x with minimal quality loss by separating the text-to-features and features-to-audio tasks and training them independently against a frozen teacher.

Paradee is a tiny text-to-speech model created by distilling a larger 82M-parameter teacher into just 8M parameters while keeping the same voice quality. The authors use a two-stage training approach: first synthesizing training data with the teacher model, then training separate text and audio components before combining them.

efficiencytrainingarchitecture
architectureefficiency

Hierarchical Continuous Diffusion Language Models

Oct 1, 2026

Hui Ren, Zihan Li, Chang Liu et al.

HC-DLM bridges discrete and continuous diffusion by coupling token generation with a shared latent trajectory, enabling better reasoning and constraint satisfaction than purely discrete or continuous approaches.

This paper proposes Hierarchical Continuous Diffusion Language Models (HC-DLM), which combines discrete token generation with continuous latent states in a single denoising process. Unlike existing approaches that treat these separately, HC-DLM uses the continuous latent as the only persistent state, reading tokens from it at each step.

architecturereasoningtraining

DMAD: Distribution Matching as Adversarial Distillation for Fast Visual Generation

Oct 1, 2026

Zhengming Yu, Junkun Yuan, Haotian Yang et al.

By framing distribution matching as a classification problem with discriminators, DMAD eliminates the memory overhead of auxiliary models while maintaining or improving generation quality—enabling practical few-step visual generation.

DMAD improves fast image and video generation by training lightweight student models to match teacher distributions without needing an auxiliary model. It uses two discriminator heads to learn density ratios directly, making the process more efficient while achieving state-of-the-art quality in one-to-four-step generation across images and videos.

efficiencytrainingarchitecture

Higher-Order Molecular Grammars for Generative and Foundation Models in Chemistry

Oct 1, 2026

Yiming Huang, Yujie Zeng, Vijay Prakash Dwivedi et al.

HGR enables sequence models to generate chemically valid molecules with perfect validity while capturing complex molecular topology, achieving top performance on generation and property prediction tasks without the computational cost of explicit higher-order encodings.

This paper introduces Higher-order Grammar Representation (HGR), a new way to represent molecules that captures complex structural features like ring systems by converting them into sequences of grammar rules. Unlike existing methods that struggle with computational overhead, HGR makes these structures compatible with standard sequence models while guaranteeing valid molecules.

architecturedataapplications

Generative Cinematographer: Composing Camera and Object Motion in 3D

Oct 1, 2026

Jiahan Zhang, Chaohao Yang, Namitha Guruprasad et al.

By lifting 2D video controls into explicit 3D space, you can resolve ambiguities in object motion and create videos where camera and object movements are geometrically consistent—a major improvement over 2D trajectory-based video control.

GenCine enables artists to control video generation by editing 3D camera paths and object motions in a scene scaffold, rather than using ambiguous 2D trajectories. The system projects these 3D controls into guidance maps that a pretrained video model learns to follow, producing videos with consistent camera-relative motion and improved geometric coherence.

multimodalarchitectureapplications

GeoLatent: Geometry-Guided Latent Structuring with Routed Optimization for 3D Reasoning

Oct 1, 2026

Yakun Zhu, Yi Bin, Yujuan Ding et al.

Separating geometry into common and residual components while routing visual learning through structured latents prevents representation collapse and improves 3D spatial reasoning from images.

This paper addresses 3D spatial reasoning from 2D images by introducing GeoLatent, which uses decomposed spatial representations (position, direction, geometry) with geometric supervision and routed optimization. The method prevents geometry collapse and ensures latents are actively used during learning, achieving state-of-the-art results on spatial reasoning benchmarks.

reasoningmultimodalarchitecture

Scaling Laws for Looped Mixture of Experts

Sep 30, 2026

Yanbei Chen, Anirudh Goyal, Raghuraman Krishnamoorthi

Looped MoE models combine two orthogonal efficiency axes: recurrence increases computational depth while sparsity expands capacity, and their scaling laws enable principled design of efficient models that match much larger dense models at the same compute budget.

This paper develops scaling laws that jointly model looped transformers (which use recurrence for computational depth) and Mixture-of-Experts (which use sparsity for capacity).

scalingefficiencyarchitecture

Looped Diffusion Transformer

Sep 30, 2026

Yong Xien Chng, Tianyi Chen, Wenwen Tong et al.

Looping computation within denoising steps is more efficient than scaling model size—you can get better images with fewer parameters by iteratively refining representations through repeated block execution.

This paper proposes Looped Diffusion Transformer, which improves text-to-image generation by repeatedly processing the same Transformer blocks within each denoising step rather than making models larger.

efficiencyarchitecturescaling

How Much of a Harness Does a Strong Agent Need for Autonomous ML Engineering?

Sep 30, 2026

Kirill Brilliantov, Alejandro Hernández-Cano, Emmanuel Abbé

For autonomous ML engineering, a minimal harness giving an LLM direct access to read, write, and bash commands performs as well as complex multi-agent systems—the model itself, not the infrastructure, drives performance.

This paper challenges the complexity of modern ML engineering agents by comparing elaborate multi-agent systems against a simple baseline where an LLM directly accesses code execution tools. The authors find that under equal time budgets, simpler agents perform as well as complex orchestrated systems, suggesting the LLM backbone matters far more than the surrounding machinery.

agentsreasoningarchitecture

Imagine3D-LLM: Teaching MLLMs to Imagine 3D Scenes Before Answering

Sep 29, 2026

Jaewoo Jung, Hyeonseo Yu, Honggyu An et al.

Training MLLMs to internally reconstruct 3D scene geometry (even in compact form) improves spatial reasoning and 3D understanding, suggesting that learning to imagine scenes is more effective than explicit geometric supervision alone.

This paper teaches multimodal AI models to reason about 3D scenes by first imagining a compact 3D representation before answering questions. Instead of relying on detailed geometric details, the model learns to assemble a coarse 3D layout from multiple viewpoints—similar to how humans understand 3D space—then uses this mental model to answer spatial reasoning questions more accurately.

multimodalreasoningarchitecture

Breakdown of Local Denoising as Semantic Speciation

Sep 29, 2026

Guangkuo Liu, Mert Okyay, Yifan F. Zhang et al.

Semantic information explains why generative models transition from local to nonlocal generation at nearly the same time—this connection becomes sharper as models scale, suggesting a fundamental phase transition in how semantic structure emerges.

This paper investigates why generative models show two key moments happening nearly simultaneously: when samples commit to a semantic class (speciation) and when local context becomes insufficient for generation (nonlocality). The authors prove these windows are causally linked through semantic information distribution, showing nonlocality must occur within speciation under natural conditions.

reasoningscalingarchitecture

Pretraining Latent Information Feedback Transformers with Teacher Supervision

Sep 29, 2026

Dor Tirosh, Ido Amos, Mor Geva

By adding recurrent feedback during pretraining via teacher-supervised state prediction, language models can improve reasoning and task performance without sacrificing training efficiency, suggesting that feed-forward architectures unnecessarily limit information flow.

This paper introduces LIFT, a transformer architecture that enables information to flow backward across layers during language model generation. Instead of the standard feed-forward design, LIFT uses teacher supervision during pretraining to train models to predict both the next token and a dense state representation derived from a teacher model.

architecturetrainingreasoning

Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE

Sep 29, 2026

Yu Xu, Yuxin Zhang, Xiao Yang et al.

Video models need different scaling strategies than language models—splitting experts by semantic role and allowing imbalanced routing produces better video generation than forcing uniform expert usage.

This paper proposes SplitMoE, a new way to scale video generation models using a split expert architecture that avoids forcing uniform expert usage.

scalingarchitectureefficiency

Telescopic Language Models

Sep 28, 2026

Zhilin Guo, Boqiao Zhang, Hakan Aktas et al.

You can train one model that works well at any depth by supervising random layer prefixes during training—no architectural tricks needed, just a smarter training objective that makes the model elastic across compute budgets.

This paper introduces Telescopic Language Models (TLMs), which train a single model that works effectively at every layer depth rather than requiring separate models for different compute budgets.

trainingefficiencyarchitecture

How to Loop MoE: Flatten the Experts, Untie the Attention

Sep 28, 2026

Shouren Wang, Chuang Ma, Mohsen Hariri et al.

To build better looped MoE models, flatten the expert hierarchy (more experts per layer, more passes) and give each pass independent attention—this lets tokens access more experts while maintaining computational efficiency.

This paper improves looped mixture-of-experts (MoE) models by flattening the expert structure and untying attention parameters. The key insight is that by doubling experts per layer and doubling passes through the network while keeping compute fixed, models can route tokens through more diverse experts, improving performance.

architectureefficiencytraining

Copy the Same, Distill the Difference: Initializing Linear Vision Transformers

Sep 28, 2026

Huaiyuan Qin, Muli Yang, Gabriel James Goenawan et al.

When adapting pre-trained vision models to use linear attention, directly copy MLP weights but distill attention behavior—this simple strategy closes the performance gap between efficient and standard transformers.

This paper shows how to initialize linear Vision Transformers (efficient attention models) using weights from standard Softmax ViTs. The key insight: copy the MLP layers directly since they learn general representations, but use distillation to transfer the attention mechanism since it's operator-specific.

efficiencyarchitecturetraining
reasoningarchitectureevaluation

Do Audio Language Models Hear and Read Distinctive Features Alike?

Sep 24, 2026

Yuanhao Chen, Peter Chin

Audio language models don't reliably represent the same phonetic features in the same direction across speech and text—suggesting these models may process the two modalities quite differently despite using a shared decoder.

This paper investigates whether audio language models represent phonetic features consistently across speech and text inputs. Using minimal pairs of phonemes that differ in single features, researchers measured whether the same distinctive features (like voicing) are encoded in the same direction in both modalities across 6 models, 7 features, and 15 languages.

multimodalevaluationarchitecture

Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs

Sep 22, 2026

Quan Nguyen-Tri, Mukul Ranjan, Zhiqiang Shen

Diffusion LLMs can be dramatically accelerated by jointly optimizing memory I/O patterns in KV caching and using the model itself for draft-and-verify decoding, rather than treating these optimizations separately.

This paper presents Flash-dLLM, a system that speeds up diffusion language models (an alternative to traditional autoregressive LLMs) by optimizing how the model stores and reuses computation results during inference.

efficiencyarchitecture

WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

Sep 21, 2026

Wangbo Yu, Kunhao Liu, Wenbo Hu et al.

By storing observations in a viewpoint-aware implicit 3D memory rather than explicit depth maps, video world models can generate longer, more consistent videos across different camera angles without running out of token budget.

WorldCrafter is a video world model that maintains consistent 3D-aware memory across different camera viewpoints and long time horizons. It uses an implicit memory system that compresses multi-view observations into tokens optimized for the requested viewpoint, enabling realistic minute-long video generation from a single image or text prompt while respecting what was seen before.

architecturemultimodalreasoning
architecturereasoningmultimodal

Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control

Sep 17, 2026

Hanchu Zhou, Brendan Lynch, Raman Goyal et al.

By treating vision and tactile signals at different timescales in a lightweight architecture, you can build fast tactile-aware robot controllers (11.9ms latency) that outperform larger models on real contact-rich tasks.

This paper presents Agile-WAM, a tactile world action model that predicts future robot states and actions for contact-rich manipulation tasks.

architecturemultimodalefficiency

dQwen3.5: Hybrid-Attention Diffusion Language Models

Sep 17, 2026

Anton Xue, Litu Rout, Aditya Akella et al.

Hybrid architectures combining attention and RNNs are surprisingly efficient starting points for diffusion language models, reaching comparable performance with 2x faster training than full-attention models.

This paper shows that hybrid-attention language models (mixing attention and RNN layers) can be efficiently adapted into diffusion language models, which generate text in any order rather than left-to-right. The dQwen3.5 models reach the same training loss in half the tokens compared to full-attention baselines, while maintaining strong performance in parallel decoding.

architecturetrainingefficiency

Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation

Sep 17, 2026

Haocheng Xi, Yiming Xie, Hexu Zhao et al.

Hybrid attention architectures that combine local Softmax with linear memory can dramatically accelerate video generation without sacrificing quality—enabling practical real-time video synthesis on modern hardware.

Video DeltaNet combines local attention with efficient linear memory to speed up video diffusion models. By mixing Softmax attention for fine details with a new Video Delta Attention mechanism for long-range context, it generates high-quality videos 14.5x faster than baseline models while maintaining visual quality.

efficiencyarchitecturemultimodal

PANORAMA: Panoptic Grounded Captioning via Mask Proposal Selection

Sep 16, 2026

Sara Pieri, Evangelos Kazakos, Shizhe Chen et al.

Combining dense captioning with pixel-level grounding requires jointly optimizing text generation and mask selection—PANORAMA shows this can be done effectively by conditioning a segmenter on phrase representations and learning which masks correspond to each phrase.

This paper introduces PANORAMA, a vision-language model that generates detailed image captions while simultaneously grounding each phrase with pixel-level segmentation masks.

multimodalevaluationarchitecture

Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environments

Sep 16, 2026

João Meneses dos Santos, Arlindo L. Oliveira

Self-reflection (validating and correcting actions at runtime) is more important than memory for making language agents reliable in interactive environments, but combining both works best.

This paper improves language agents for interactive tasks by adding two cognitive modules to SwiftSage: a memory system that stores and retrieves important experiences, and a self-reflection system that validates actions and fixes mistakes. Testing on ScienceWorld shows the full system performs best, with self-reflection being the most critical component for handling long-horizon tasks.

agentsreasoningarchitecture

How Model Growth, Recursion, and Boundary Operators Influence Scaling Exponents

Sep 16, 2026

Zixi Chen, Akshay Vegesna, Samip Dahal et al.

Architectural design choices like recursive depth and boundary operators can fundamentally change scaling laws, enabling exponential compute efficiency improvements—challenging the idea that scaling laws are fixed.

This paper shows that architectural changes—specifically model growth through recursive depth (looping) and boundary operators—can improve how efficiently transformers use compute during training. A 7.4B looping model matches GPT-3 13B's performance with 20× less computation, with efficiency gains that grow larger at bigger scales.

architecturescalingefficiency

LACE: Layer-Wise Compression for Dynamic Frame Rate Codecs

Sep 15, 2026

Thanapat Trachu, Samuele Cornell, William Chen et al.

Layer-specific compression in neural codecs reduces computational overhead by letting different quantization layers compress at their own optimal rates, improving efficiency without sacrificing quality.

This paper introduces LACE, a neural audio codec that compresses speech at different rates for each quantization layer rather than using a single compression step. By allowing each layer to have its own segmentation boundaries, LACE reduces sequence length and computational cost while maintaining audio quality.

efficiencyarchitecturetraining
reasoning
architecture

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

Sep 8, 2026

Siting Li, Zhengyang Wang, Simon Shaolei Du et al.

When building multimodal models, image tokenizer choice matters not just for image quality but for how well it helps the model learn text and vision together—and the best tokenizer for one task may not be best for another.

This paper studies how image tokenizers work in multimodal AI models by building a controlled training setup that tracks how well different tokenizers perform across text, image generation, and image understanding tasks.

multimodaltrainingarchitecture

DeCAL: Towards Physically-Grounded Dexterous Vision-Language-Action Models via Contact-Aware Latent Co-Imagination

Sep 8, 2026

Yankai Fu, Ning Chen, Junkai Zhao et al.

Adaptive tactile fusion and joint visual-tactile prediction enable VLA models to handle contact-rich manipulation where vision alone fails, achieving 71% success on dexterous tasks.

DeCAL is a vision-language-action model for robot hands that combines visual and tactile (touch) sensing to perform complex manipulation tasks. It uses specialized AI experts working together to understand scenes, imagine future states, and generate actions—all while explicitly modeling contact dynamics that cause visual occlusions during dexterous manipulation.

multimodalagentsarchitecture
multimodalarchitectureapplications

Design Docs Are All You Need: An AI-native Machine-Learning Performance Tool

Sep 4, 2026

Samuel Kushnir, Kimia Noorbakhsh, Kavya Sreedhar et al.

Design documents with worked examples can be more maintainable than code for ML performance tools, since AI agents can reliably regenerate implementations from them, reducing the cost of keeping performance models up-to-date with new hardware and models.

SMART is a machine-learning performance modeling tool that replaces traditional code with natural-language design documents. AI agents regenerate the entire implementation from these docs on each update, eliminating tech debt while maintaining accuracy—validated against real systems like DeepSeek-V3 on TPUs.

efficiencyarchitectureapplications

Embedded Graph Flows for Categorical Graph Generation

Sep 4, 2026

Ethan Ma, Zihan Wang, Chris Siu Yeung Chow et al.

Learning continuous embeddings for graph categories instead of using fixed one-hot vectors improves molecular graph generation quality, as shown by EGF's superior performance on standard benchmarks.

This paper introduces Embedded Graph Flows (EGF), a generative model that creates realistic molecular graphs by learning continuous embeddings for node and edge types instead of using fixed one-hot vectors. The model uses a permutation-equivariant transformer to gradually transform noise into valid graph structures, achieving state-of-the-art results on molecular benchmarks like QM9 and ZINC250k.

architectureevaluation

LLM-Driven Algorithm Design for Quantum Circuit Synthesis based on Binary Decision Diagrams

Sep 4, 2026

Yoonju Sim, Federico Berto, Chuanbo Hua et al.

LLMs can discover domain-specific algorithm improvements by searching over heuristic families rather than predicting solutions directly—here achieving 71% win rate on quantum circuit optimization by aligning variable ordering with actual quantum cost rather than proxy metrics.

This paper uses large language models to discover better algorithms for ordering variables in quantum circuit design. The key challenge is that quantum circuits implementing Boolean functions need optimal variable orderings to minimize quantum cost, but existing heuristics optimize for the wrong metric.

reasoningarchitecture

How Does mHC Use Its Residual Streams? Selective Routing and Near-Identity Mixing

Sep 4, 2026

Pengxiang Zhao, Xing Li, Xianzhi Yu et al.

Multi-stream architectures like mHC provide more capacity than models actually use—most layers rely on just two streams, and late-layer stream mixing can be removed with minimal performance loss, suggesting opportunities for efficiency improvements.

This paper investigates how DeepSeek-V4-Flash uses its multi-stream residual pathways (mHC architecture). Researchers found that despite having four parallel streams, the model typically uses only two streams per layer, with minimal mixing between streams in later layers. Interventions show that early-layer mixing is critical for performance, while late-layer mixing is largely redundant.

architectureefficiencyevaluation

Para-Pipe: Exploiting Hierarchical Operator Parallelism of ML Computational Graphs on SoCs

Sep 3, 2026

Yujie Zhang, Huiying Lan, Ehsan Aghapour et al.

For edge AI deployment, combining hierarchical operator parallelism with pipelining can achieve better energy efficiency and latency than using either strategy alone—Para-Pipe shows 11-23% energy improvements on real SoCs.

Para-Pipe is a framework that optimizes deep learning inference on edge devices (SoCs) by combining pipelining and parallel execution of neural network operations.

efficiencyarchitecture

The Natural Language Interaction Protocol and Standard for AI Agents

Sep 3, 2026

Luyi Xing, Rasit Onur Topaloglu, Ranjan Sinha et al.

NLIP is a practical interoperability standard for AI agents that lets you build agents independently and have them work together seamlessly, similar to how different web services communicate via HTTP.

This paper introduces NLIP, a standardized communication protocol that lets AI agents built with different frameworks and tools talk to each other. Like HTTP for web servers, NLIP provides a common language for agents to exchange messages, coordinate work, and access shared tools—enabling organizations to mix-and-match agent systems without rebuilding everything from scratch.

agentsarchitectureapplications