Research Papers Kimi K3 is an open Mixture-of-Experts model with 2.8T parameters (104B activated per token), native vision input and a 1-million-token context window, built on Kimi Delta Attention and post-trained with reinforcement learning across general, agentic and coding tasks.
FreeToken rethinks local inference around two realities of edge AI: agent workloads keep changing shape, and every machine's resource balance differs. It co-designs model layout, expert residency, CPU-GPU execution and memory management instead of committing to a fixed offloading strategy.
Prime Agent is an open harness for long-horizon coding agents built on the Recursive Language Model idea: a persistent IPython session does test-time computation, while histories, memories, skills and subagent specs carry over between runs.
After Meta's closed Muse demo showed what a single-person agent with lasting memory could look like, this report reconstructs how Muse is built from public sources, then ships nanoMuse as its GPL-3.0 open counterpart: one agent across the devices you own, sharing a single conversation over a relay anyone can self-host. Every action passes a Sentinel gate before it executes, memory is stored as plain files the owner can read, and the underlying model is the user's choice.
MOSS-Image is an open generation system aimed at the gap where open-weight image models still struggle with dense text and intricate layouts. It pairs a 5B single-stream diffusion transformer pretrained from scratch, which the authors say enables bilingual long-form text rendering and precise layout control, with an intention-aware agentic prompt enhancement module that turns loose user requests into detailed content and layout specifications. The report introduces two new benchmarks, LongLongTextBench and RealLayoutArena, and reports that MOSS-Image is state of the art among open-weight models on long-form English and Chinese text rendering with 96.3% average text coverage, and ranks second in end-to-end layout-rich generation. The authors say they will release the model, training recipe, and evaluation framework.
Symbolic music models plan composition explicitly but never sound like a finished recording, while audio models produce full songs that keep the composition hidden. YuE2 combines both in a single AR-NAR Mixture-of-Transformers that plans symbolically, expands the plan into semantic music tokens, and renders the full song, using two new representation models (MERT2, SheetSage2) to learn from recordings that lack aligned scores. The authors report a SongBench Global Avg of 6.73 on WildSongBench (6.96 at best-of-8), and in expert listening the best-of-8 samples were preferred over Suno v4.5 and roughly tied with Suno v5; the readable score also makes the output editable by external language-model agents.
Raven argues that hand-crafted agent harnesses are too domain-coupled to scale with long-horizon, cross-domain workloads. It treats each model–harness pair as a composable unit of intelligence: a Host Agent decomposes goals and coordinates specialized sub-agents, a host archive and EverOS retain experience across tasks, and Skill Forge turns that experience into reusable procedures — with theory on expanding reliable task coverage under one resource budget.
OpenGCUA rebuilds the computer-use agent stack in the open, motivated by the observation that frontier systems like GPT-6 Astra's computer use are closed. Its core idea is to let the model choose between visual GUI actions and direct code execution, so intermediate results can steer later steps; it also proposes harness-to-model decision transfer, moving selected observation and task-state decisions out of fixed harness rules and into the model itself. The authors report that on OSWorld-v2 their full configuration improves task scores across all eight evaluated models (including GPT-6 Astra and Claude Opus 5), raising mean partial score by +19.6 points and cutting reported steps per task by 53.0% on average versus batched GUI-only execution. Alongside the implementation and prompts they release evaluation trajectories from all eight models, plus early versions of a real-device harness and a cross-device SDK covering computers, mobiles, and browsers; on four real macOS tasks their harness shows selected Qwen3.8-Max runs use 19.4 times as many reported reasoning tokens as GPT-6 Astra, which they highlight as a gap in real-device execution efficiency.
Most video benchmarks score a single prompt-to-clip pipeline, but streaming models must keep generating while new instructions arrive mid-stream. StreamEval scripts 300 interactive test cases (15-300s each) across five tasks: avatar dialogue, avatar performance, world generation, instruction editing, and reference editing, where multimodal requests arrive incrementally and each case demands both state changes and content preservation. Scoring 37 models on 10 metrics across four dimensions, the authors report that strong visual quality does not translate into following instructions: the best request-execution scores are only 36.90 (avatar performance Following), 25.93 (world generation Following), and 15.00 (reference editing Edit Success) out of 100, flagging reliable real-time request handling as the field's main open problem. Leaderboard at streameval.ai.
VoiceMem gives duplex speech models a working memory that keeps facts and emotional context across a conversation, so a voice assistant can remember both what you said and how you said it, in real time.
This paper evaluates GPT-6-Astra as a general embodied agent rather than a robot-specific model, running it across 12 benchmarks covering tabletop manipulation, humanoids, mobile platforms, and dexterous hands, plus 8 real robot setups. The authors find that with in-context demonstrations and longer execution horizons it shows broad competence, but report that a System 2 reasoner paired with a low-level controller is expensive, that demonstration collection is costly, and that failures often come from irreversible physical mistakes rather than reasoning. Their conclusion: GPT-6-Astra has largely crossed the threshold of embodied competence, but not yet practical embodied intelligence, and they sketch a harder next-generation benchmark built from real needs.
MemT-1.0 is a family of memory foundation models (Nano and Flash) that make memory an internal computation step instead of stuffing past context into the prompt: built on the Metis architecture with Gated DeltaNet-2 and a Mixture-of-Storage layer, the model updates token-level memory within each turn. The authors report it beats parametric-memory baselines on average, stays competitive with retrieval-augmented methods, and at 128K-token history it achieves about 1.86x end-to-end inference speedup with 66x less persistent storage than keeping the full KV cache.
Edge0 is an open-source inference engine that serves a 35B mixture-of-experts model from a single 24GB GPU by loading expert weights from SSD on demand; a trained per-layer prerouter predicts the next layer's expert choices one token ahead so SSD reads overlap with compute, and a recovery LoRA pays back the quality lost to int4 quantization. The authors report 20 tok/s at about 3GiB of peak active memory, within a few points of the fp16 teacher on five public benchmarks.
This 66-page position-style paper formalizes harness self-evolution: an agent keeps its model weights fixed but continually revises its own software scaffolding, including prompts, memory, workflows, executable components, and plugins. Accepted revisions persist across runs, and the authors separate what can be updated from the proposal and admission mechanisms, analyzing reliability with PAC-learning-style bounds and an admission rule that requires fresh independent validation data. The paper argues this demands continual validation plus independent safeguards on runtime privileges and consequential actions.
Frontis-MA1 ships with OpenMLE, an open full-stack system (task gyms, RL operator learning, evolutionary search) for studying recursive self-improvement. Its 35B meta-evolution agent, post-trained around Draft, Improve, Debug and Crossover program-evolution operators, claims MLE-Bench Lite results the authors say exceed GPT-5.5 plus Codex, with components transferring to held-out NatureBench Lite.
HexaShell is a deliberately thin robot-control harness: it only converts a coding agent's commands into motion, giving the agent three calibrated camera views, one command line, failure logs, and the ability to write reusable position-estimation tools based on camera-ray triangulation and ray-plane intersections. Trained on 10 layouts with one seed per task and then frozen, the tool library is tested on 10 unseen layouts per task across all 42 RoboDojo tasks using Codex with GPT-6-Astra at medium reasoning effort. The authors report 39.7% success and a 48.0 progress score, which they call a new state of the art on RoboDojo, with pushing at 92.5%, memory tasks at 65.0%, but tight insertion only 8.3%. Their analysis says the bottleneck is physical contact rather than reasoning: the agent often knows what to do but cannot execute the motion reliably. The tools are open-sourced.
Scaling laws are well established for language models in the physical world, but do they hold for embodied agents acting in virtual worlds with different graphics, physics, controls, and objectives? This paper runs a controlled study of continued pre-training across three axes, with 120 configurations spanning four model sizes (0.8B to 27B), six data budgets (50 to 1,600 hours), and five data mixtures from Minecraft-only to diverse multi-world corpora. The findings: more data helps within the training worlds, data diversity transfers to related worlds but barely to distant ones, and larger models use data more efficiently yet generalize worse to distant worlds under single-world training. Training also splits into an early interface phase with broad cross-world gains and a later world phase where gains narrow to familiar worlds. The authors fit an empirical scaling law relating validation loss to volume, diversity, and size, and note that persistent gaps in distant worlds mean generalist embodied agents will need training stages beyond continued pre-training.
Making a long film with a video-generation agent breaks down at the seams: passing prompts or a few recent frames between shots gives the model neither a shared spatial frame for revisiting a scene nor a time-indexed record for replaying an earlier event. ATAS compiles a screenplay into shot requests with explicit location, camera, and event-time references, and keeps two memories behind the generation loop: a scene memory of registered static geometry and an event memory of dynamic observations indexed by scene, event, and time. Each generated clip is reconstructed and registered back, admitted only after review and state checks. The authors report that this improves all measured cross-shot consistency metrics across four video generators, outperforms compared agents and 3D-guided methods on spatial revisits and event recapture, and earns the highest human ratings on all six evaluated criteria.
Fine-grained perception usually means stitching together separate specialist models for detection, segmentation, counting, and OCR. WeVisOne-VL (from the WeChat Vision team) unifies object-level and pixel-level tasks in one VLM: a lightweight pixel decoder plugged in via a special <OBJ> token lets the same model output boxes, points, masks, and even alpha mattes at whatever granularity the user asks for. Trained with three-stage curriculum SFT plus mixed-task RL, the authors report leading results across 30+ benchmarks while still refusing queries where the requested object is absent, a failure mode they explicitly target for real-world deployment.
This survey organizes in-context learning for robotics around one question: how can a robot pick up a new task from demonstrations or interaction while its network weights stay fixed at deployment? It groups the literature into four families based on how context connects to execution: context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill- or agent-based execution, then compares their transfer assumptions and how training, correspondence and memory make the context useful. It also reviews evaluation practices that separate responsiveness to teaching, physical transfer, and gains from retained experience, and closes with a research agenda toward physical recursive self-improvement. A companion GitHub repo collects the papers.