Browse released weekly reports in reverse chronological order.
This week's evidence treats agent memory as a problem of selecting, representing, and maintaining useful experience, not simply extending context windows.
The representative studies evaluate retrieval as a component of agent reasoning, exposing mismatches between generated queries, retriever capabilities, and evidence use that final-answer scores alone do not reveal.
This week's theme links language-model evaluation with data efficiency, multilingual coverage, compression, and instruction following.
The representative research broadens LLM evaluation beyond single-turn correctness to multi-task interaction, learning across repeated attempts, and reward quality under increased inference compute.
This week's theme links socially appropriate AI participation, cooperative effects on human partners, and physically feasible embodied behavior, framing evaluation as more than response quality.
The theme centers on evaluating whether language-model instructions, reasoning, and communication translate into effective embodied navigation rather than merely plausible text.
The theme centers on preserving useful visual representations while learning new classes.
This week's theme links physically grounded demonstration generation with reinforcement-learning fine-tuning to address scarce robot data and limited demonstration coverage.
This week's evidence emphasizes evaluation beyond aggregate generation quality: realistic editing requires interpreting context, making precise changes, and preserving unaffected content.
Work on autoregressive video models centers on two linked problems: preserving useful visual history over long rollouts and learning representations that keep generation efficient and coherent.
The theme centers on open, multi-stage LLM post-training in which evaluation and verification guide data selection, preference tuning, and reinforcement learning.
The theme centers on adapting representations to heterogeneous visual and linguistic inputs rather than relying on uniform processing.
This theme centers on making long-context LLM inference practical by reducing the rapidly growing KV cache memory footprint.
This theme centers on moving beyond direct answer generation toward models that simulate underlying system dynamics before making decisions or predictions.
This theme centers on evaluating speech-language systems that listen and speak concurrently, moving beyond conventional turn-based assessment.
This week's theme centers on how multimodal models should represent visual information when a single system is expected to understand, generate, and edit images.
This theme covers speech recognition systems designed to operate beyond clean single-speaker conditions, addressing noisy or soundproof environments, overlapping speakers, and channel mismatch.
This theme centers on how learning systems can improve themselves or revise their own reasoning without silently accumulating harmful changes.
This week's theme centers on evaluation shifting toward more realistic, high-constraint benchmarks rather than simplified test settings.
This theme centers on making 3D/4D reconstruction and rendering more practical under sparse-view or monocular conditions by moving beyond purely photometric fitting.
This theme centers on rethinking what reinforcement learning should optimize, beyond standard expected-return training.
This theme focuses on evaluating whether language models and multimodal systems can genuinely reason across visual, linguistic, and spatial representations rather than rely on superficial benchmark signals.
This theme centers on models that unify multiple modalities while making their representations more controllable and more meaningfully testable.
This theme centers on the shift from evaluating agents in fixed setups to scaling the environments, interaction loops, and runtime scaffolds that shape agent behavior.
This week's theme centers on moving model evaluation beyond narrow offline metrics toward executable, task-grounded benchmarks for embodied world models and multimodal GUI agents.
This theme concerns making visual generation and reconstruction explicitly grounded in geometry, camera parameters, and scene structure.
This week's reinforcement learning work focuses on making RL systems more general, computationally efficient, and stable under demanding training regimes.
This theme centers on how foundation models are being evaluated and improved for scientific and multimodal use, with emphasis on pretraining strategy, data realism, and efficient cross-modal design.
This week's theme centers on making neural surface reconstruction both generalizable and efficient under sparse-view conditions.
This theme tracks the shift from static, centrally orchestrated LLM agents toward self-evolving, decentralized multi-agent systems that adapt their roles, memory, tools, and coordination over time.
Benchmarking work is increasingly targeting harder, more credible evaluation settings for multimodal question answering, especially where models must retrieve or ground evidence rather than rely on shallow multiple-choice cues.
Evaluation for computer-use and GUI agents is shifting from task success measurement toward diagnosis of safety, efficiency, and cross-platform capability.
Representative papers converge on a shared limitation of static imitation and standard model-based control: policies struggle with distribution shift, sparse rewards, and weak exploration once they leave the expert data regime.
This theme centers on how increasingly capable language and multimodal models should be evaluated and controlled as they move beyond text-only benchmarks.
This week's theme centers on evaluating and improving embodied manipulation in contact-rich settings where vision-only supervision and narrowly collected robot data are insufficient.
This theme centers on evaluating AI coding assistants and software-engineering agents beyond narrow benchmarks or final-output metrics.
This theme centers on evaluating LLM-based coding agents on realistic software-engineering tasks, especially repository-level issue resolution.
This week's theme centers on using embodied world models not just to generate realistic futures, but to evaluate, supervise, and improve robotic policies.
This week's theme centers on evaluating LLM behavior in interactive, socially grounded settings rather than judging single-turn text quality alone.
This theme centers on evaluating how safety alignment breaks down when language models operate across images and broader multimodal inputs.
This theme centers on making visual generation and reconstruction more practical by replacing slow per-scene optimization with feed-forward Gaussian representations and by improving video diffusion efficiency.
This week's benchmarking work in affective AI addresses the need for evaluation resources that better reflect the complexity of human emotion.
This theme centers on scientific reasoning as a demanding evaluation setting for LLMs, where models are tested on expert-level, multi-step problems rather than surface-form recall.
This week's work centers on robot manipulation systems that move beyond reactive control by integrating predictive world models and richer contact feedback.
This week's work on video diffusion models advances along three coupled fronts: generation quality, temporal/spatial consistency, and inference efficiency.
This theme centers on reconstructing and rendering dynamic 3D/4D scenes from monocular or casual video using Gaussian-based representations.
This week's theme centers on how vision-language models should be evaluated and improved when standard web-scale data and static benchmarks fail to capture real capability.
This theme centers on evaluating whether LLM systems can answer long-form, multi-document questions while staying grounded in evidence drawn from both text and visual materials.
This week's theme centers on evaluating and improving LLM agents through richer environments rather than static datasets alone.
This week's theme centers on making learned representations interoperable across tokenizers, models, and modalities.
This theme centers on reframing segmentation from conventional discriminative per-pixel prediction toward generative mask construction and refinement using pretrained diffusion models.
This theme centers on making reasoning-oriented LLMs more efficient in both training and inference, rather than treating stronger reasoning as purely a scaling problem.
This week's theme centers on how video models are being evaluated and redesigned for stronger temporal reasoning, especially in action and egocentric settings.
This theme centers on new benchmarks and evaluation frameworks for instruction-based image editing, motivated by the gap between advancing visual generation and reliable edit assessment.
This week's theme centers on evaluating and improving LLM-based research and problem-solving agents beyond one-shot task success.
This week's papers advance world modeling away from monolithic black-box predictors toward structured, modular architectures designed to better capture the dynamics of diverse environments.
This week's theme centers on making model merging more controllable, scalable, and robust as the number of fine-tuned expert models grows.
This week's work marks a shift from evaluating multimodal models on static perception toward testing whether they can form actionable, physically grounded world models.
This week's AI safety research emphasizes the shift from broad concern about AI harms toward structured governance and quantitative risk-modeling frameworks.
This theme centers on evaluating and structuring LLM reasoning in settings where static prompting or generic inference heuristics break down—especially when retrieval, domain knowledge, and multi-step decision rules must interact.
This week's theme centers on applying reinforcement learning to move recommendation beyond greedy next-item prediction toward long-term user engagement.
This week's papers treat representation quality and cross-scale alignment as a central bottleneck in both generative modeling and general visual pretraining.
This week's theme centers on how vision-language and embodied models are being tested and redesigned for navigation when spatial reasoning, long-horizon decision-making, and safety become bottlenecks.
This week's theme centers on how LLM-based research agents should be assessed and scaffolded as they move beyond writing support into research planning, experimentation, review, and publication workflows.
This week's theme centers on equipping vision-language models with explicit geometric and navigational structure for embodied tasks, moving beyond brittle prompting or task-specific heads.
This theme centers on diffusion models that move beyond generic text-to-image generation toward more structured, grounded, and computationally practical image editing and perception.
This week saw continued progress toward unified models that combine image generation, editing, and understanding within single autoregressive or hybrid autoregressive-diffusion architectures.
This theme centers on coordinating multiple LLM-based agents to handle tasks beyond what a single model instance can easily support.
This week's theme centers on methods that recover richer scene structure and semantics from limited video observations.
This week's theme centers on benchmark work that evaluates world, video, and multi-view generation models beyond surface-level visual quality.
This week's reinforcement learning theme centers on making agents learn richer behaviors through curriculum design and modular skill representations.
This theme tracks activation steering as an inference-time method to control and adapt language models without modifying parameters.
This week's papers frame advanced video and multimodal generative systems as emerging world models rather than mere content generators.
This theme centers on how to evaluate LLM-based agents for scientific research and complex information seeking under realistic, controlled conditions.
This week's theme centers on discrete and masked diffusion language models as an alternative to autoregressive LLMs, with particular emphasis on how decoding order shapes capability and efficiency.
This week's evaluation work highlights persistent gaps between how visual models are assessed and the conditions they face in practice.
This theme addresses how to evaluate and improve models' understanding of temporal structure in video.
This week's work reflects a shift from building GUI-capable VLM/LLM agents toward evaluating them more rigorously across platforms, capability levels, and failure modes.
This week saw multiple new competition benchmarks that extend image restoration evaluation beyond single-degradation settings.
This week's representative papers address how to scale large language models more efficiently through mixture-of-experts architectures and smarter pre-training data-mixture design.
This week's progress centers on making diffusion-based multimedia generation more temporally coherent and controllable as these models expand from images into video and audio.
This week's theme centers on evaluating 3D reconstruction under realistic adverse conditions—noisy video, human-object interaction, and sparse or degraded observations.
This week saw continued progress on using transformer-based pretraining to enable in-context adaptation in sequential decision-making without weight updates.
This week's papers focus on making LLM agents more reliable on complex, long-horizon tasks by improving how they store, extract, share, and secure knowledge.
This week's representative papers highlight that medical AI progress hinges on clearer evaluation frameworks and richer clinical context, not only on stronger models.
This week's papers center on how to organize LLM-based multi-agent systems for complex, real-world tasks.
This theme centers on how LLM outputs can be attributed to supporting documents so that generated answers are more transparent, verifiable, and trustworthy.
This week's papers focus on making multimodal foundation models more efficient without sacrificing broad utility.
This week's theme concerns adapting and evaluating speech models when labeled in-domain data are scarce, domains shift, or speech departs from typical patterns.
This week's papers frame AI deployment as an environmental and governance challenge.
This week's evaluation work pushes beyond narrow benchmark settings toward broader tests for LLM- and VLM-based agents.
This week's theme centers on privacy evaluation in federated learning, where shared gradients, parameters, or predictions can leak sensitive information even when raw data stays on-device.
This week's papers treat the environmental impact of AI infrastructure as a direct evaluation concern.