2026-10-13

Where Muse Spark aimed to be a lightweight model that could perform at low latency, the Wonder models are larger and said to be more capable, with higher benchmark scores and better results on complex tasks like coding.

Decoder hidden states from different layers of large language models (LLMs) often encode complementary information. However, existing speculative decoding methods typically rely on a small set of target-model layers chosen using fixed, position-based heuristics. Selecting an optimal subset of layers is inherently a combinatorial problem, making exhaustive evaluation infeasible even for moderately deep models. In this work, we present a principled framework for multi-layer hidden-state selection

The evaluation of coding agents generally strives to be as realistic as possible. In our study, we collect 4,782 agent sessions of real software engineers in JetBrains IDEs, which we call Production Sessions. Since our subject is interactive agents, we study the sessions with at least three user messages (33% of the sample). These long sessions differ from issue-derived benchmark tasks in two ways: (i) user requests span a far wider mix of task types - questions about the project's code, plannin

We're launching Precision Numerics by releasing Boole-1.7B, a 1-bit version of Qwen3-1.7B. 247 MB packed weights: 13.9× smaller than BF16. 85.9% of the BF16 model’s average score across our 10-task evaluation - outperforms 1-bit Bonsai-1.7B and nearly matches 1.58-bit Bonsai-1.7B. Try it locally in your browser (it's for low-latency tasks using English and non-thinking mode), or use it with vLLM or llama.cpp. We'll ship more capable models for the open-source community. Let us know what you what

Continual post-training faces a dual challenge of acquisition and retention: models must learn new capabilities while preserving prior knowledge. While Self-Distillation Fine-Tuning (SDFT) mitigates this using self-supervision with in-context demonstrations, it relies on strong in-context learning and instruction-following abilities. Consequently, SDFT fails on smaller models critical for low-latency, on-device use. To enable continual learning at smaller scales, we first identify the root cause

The models add "active memory" and "active focus" abilities.

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now New benchmark shows how AI models lack intuitive visual reasoning:…Humanity’s Sixth Sense highlights some shortcomings of current AI systems…AI startups Elorian and Scale AI have built […]

Process reward models (PRMs) have become a key component for LLMs, as their step-level feedback supports both post-training and test-time reasoning. However, training strong PRMs remains costly: human step annotation is difficult to scale, while Monte Carlo estimation is computationally expensive and can drift from the intrinsic correctness of steps. To get effective PRMs at low cost, we introduce TIPS (Thinking-Induced Process Supervision), an outcome-only reinforcement learning (RL) framework

arXiv:2610.12813v1 Announce Type: new Abstract: LLM agents increasingly operate over sensitive resources, such as private documents and mutable system states, which can create serious safety and security risks. Runtime guardrails should therefore make inspectable, reproducible, and deterministic allow or refuse decisions before potentially irreversible agent actions are executed. However, safety and security requirements are rarely complete before deployment, so guardrails need to keep adapting

arXiv:2610.12598v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) provides reliable supervision for software engineering agents through executable tests, but binary rewards fail to distinguish trajectories with substantially different debugging quality. We investigate whether the current policy can judge its own trajectories and use these judgments to improve task performance without a separately trained reward model. To this end, we introduce Behavioral Self-

On-policy distillation (OPD) has become increasingly popular for transferring teacher capabilities to student models. In this work, we ask a critical research question: Is on-policy sampling always beneficial for distilling arbitrary teacher-student pairs? We show that a simple alternative, Semi-OPD, which distills from offline rollouts generated by the initial student, can often outperform OPD in both accuracy and training efficiency. Across 17 teacher-student pairs ranging from 1.5B to 235B pa

arXiv:2610.12822v1 Announce Type: new Abstract: Multi-agent systems are increasingly used to solve coding tasks by coordinating agents that share code and plans. These agents are often optimized with reinforcement learning (RL) using automated evaluators, but bypassing evaluation can earn reward without producing a genuine solution. Prior work has largely treated reward hacking as a single-agent failure. In this work, we show that a reward-hacking strategy elicited in one agent can propagate to

arXiv:2610.12655v1 Announce Type: new Abstract: LLM agents can execute increasingly complex tasks, but effective assistance also requires aligning their actions with goals and constraints that users may leave unstated. We study this problem as interactive intent alignment through clarification. Yet clarification is useful only when the agent can turn the answer into better decisions. As training changes how the agent handles incomplete information and uses additional answers, the same clarificat

Language model agents increasingly operate over long interaction streams, requiring online adaptation to distribution shifts through continual learning. However, existing agent harnesses typically either evolve their states only through forward information flow or rely on additional replay or validation data for harness optimization. To address this, we propose ReHarness, a recurrent harness framework for agentic continual learning that complements forward read and write with backward revision u

Recent GPU kernel optimization agents retain what they learn in knowledge bases or as distilled skills. Kernel benchmarks score each attempt's implementation for correctness and speed but leave the reuse value of retained experience unmeasured. We introduce MetaKernelBench, which measures whether experience distilled from an attempt in one kernel domain-specific language (DSL) improves a fresh attempt at the same problem in another. Its 74 problems are fused subgraphs in six families, each posed

arXiv:2610.12789v1 Announce Type: new Abstract: Self-evolving LLM agents typically improve how they act through accumulated experience, reflection, or skills. Yet even procedurally capable agents can fail when their beliefs about environment dynamics are incomplete, incorrect, or outdated. We study a complementary axis of self-evolution: evolving those beliefs during deployment. Such beliefs can support multiple tasks, but incorrect ones can also propagate across them. We introduce EvoBelief, a

arXiv:2610.12593v1 Announce Type: new Abstract: The continued improvement of large language models increasingly depends on synthetic experience, yet effective data synthesis for interactive agents requires not only executable and verifiable trajectories, but also interactions that provide learning value for the current task agent. Existing methods increasingly incorporate environment feedback, data validation, or learner failures, but still face a feedback-to-decision gap: observed weaknesses do

Video understanding agents acquire evidence through an executable harness that controls what they observe and how they use those observations. However, execution traces contain only the evidence acquired by the current harness, leaving competing explanations for failure unresolved and limiting the basis for self-improvement. We introduce Video-RSI, a framework for recursive self-improvement in which a video understanding agent uses its own language model to revise its harness. Through active vid

gguf's: " MiniCPM-V 4.7 is the latest model in the MiniCPM-V series, available in a 35B-A3B variant, with 35B-A3B being the first MoE model in the MiniCPM-V series . It builds on the family's single-image, multi-image, and video understanding capabilities, combining stronger foundation capabilities with efficient visual encoding and on-device inference. Notable features of MiniCPM-V 4.7 include: 📱 The First Open-Source MoE Model with Lossless, Smooth On-Device Inference. With our Lookahead Exper

The self-learning coding agent for your terminal Discussion | Link

Hey everyone, I made whatmodel.sh To use it, Enter your RAM and GPU See what models you can run locally with its rank and the command to run it Compare them with API models on the same score Building it made me realise a 32GB Mac can now run a model that scores on-par with GPT-5, completely locally and for free (e.g. Qwen3.8 27B at 4-bit). Each model gets one score fitted across 19,731 published benchmark results, and every result links to its source. The method is at whatmodel.sh/method , and t

Visual representations are largely developed for understanding, yet what makes them useful for generation remains unclear. We study this question in pixel diffusion transformers and uncover two key insights. First, better representation quality does not necessarily translate into greater generative utility, as externally aligned semantics can be progressively lost during pixel synthesis. Second, architectural separation with specialized pixel decoders mainly redistributes semantic and spatial mo

arXiv:2610.12659v1 Announce Type: new Abstract: Data-centric scaling of RLVR has focused on collecting or generating verifiable training problems. However, corpus size alone does not determine whether the dataset provides a useful learning signal for a target policy. To guide the synthesis of effective RLVR data, we experimentally characterize dataset effectiveness as a function of its problems, the target policy, and the training rollout budget. Under this characterization, we show that a large

arXiv:2610.12596v1 Announce Type: new Abstract: Agentic large language models now tackle complex tasks, yet strategic reasoning over long horizons is under-studied: most agentic benchmarks are short and episodic, and performance on long tasks is hard to dissect into stored knowledge, reasoning and luck. We introduce SpireBench, in which models play the complete commercial deck-builder Slay the Spire (StS), where early decisions decide long-term success. Across 20 large language models (LLMs) at

arXiv:2610.12536v1 Announce Type: new Abstract: EliteTable is a self-improving program-search algorithm guided by large language models (LLMs) that uses a fixed-capacity table of high-scoring solutions to guide new proposals, edits of existing solutions, and remixes of several solutions. Performance was evaluated across six Gymnasium control tasks, each with ten generations and fifty candidate slots per generation. Each task was evaluated across ten independently initialized searches. Candidate-

Release: datasette-ports 0.4 My plugin for detecting and listing all Datasette instances on the current machine now works with the latest version, and displays the detected PID as well. Knowing the PID makes it easier to attach the new Python 3.15 profiler . Tags: datasette

Foundation models can serve as clinical agents through tool-use harnesses. However, conventional medical benchmarks assess reasoning over preselected evidence rather than the ability to seek it across clinical records and longitudinal imaging. We propose CASE: a series of role-specific Clinical Agents for Seeking Evidence, together with a tool-use harness and an agentic post-training framework for compact vision-language policy models. We further introduce a longitudinal multimodal benchmark bui

Attention Residuals (AttnRes) enables content-dependent retrieval over network depth, which can be viewed as single-stream softmax attention along the depth axis. Inspired by the success of multi-head attention along the sequence axis, we introduce Multi-Stream Attention Residuals (MSAR), an extension of AttnRes that generalizes this mechanism to multi-head depth attention, yielding substantial performance gains with minimal additional overhead. MSAR draws stream diversity from sequence context

Russian bloggers warn Russia’s hyper-centralized digital economy is vulnerable.

Hi HN! I'm Saarth. I'm in 9th grade, and I built Atticus. Atticus is a fast and cheap Mac assistant that can do tasks for you. You hold down Control-Option-Space, speak or type your task, be it complex or a simple browser search, and Atticus will get on it. The interesting part is that simple tasks, like opening apps, playing a Spotify playlist, or web queries, don't touch AI, and instead are processed locally and typically finish in under a second. For example, prompting Atticus to open apps ta

In August, I posted here about TeXbrain, a LaTeX editor that runs completely in the browser, with no need for an account or server. (I didn't excpect it to get that much feedback, bug reports and help from here!) This is the same idea for the other half of my studies, the handwritten notes. I take them on a convertible laptop, and I've had a hard time with apps for that, taking notes from a device to another, exporting as PDF, needing an account for many of them, etc. So I've been working on the

游戏打到21:0，AI甚至不用再调用大模型了

arXiv:2610.12842v1 Announce Type: new Abstract: Neural networks routinely develop representations in which individual neurons contribute to multiple distinct functions, leading to the widespread view that such networks lack modular organization. We challenge this conclusion by introducing a formally nested hierarchy of modularity: unit-level, which requires distinct computations to occupy disjoint neurons; orthogonal, which asks whether a rotation reveals functionally independent subspaces; and

Humans perceive far more in a scene than what is explicitly depicted: a single glance captures past causes and future trajectories; a quick peek determines if a vehicle can fit between two parked cars; a few seconds of video reveals who holds authority in a room; and a fleeting clip highlights subtle abstract patterns like unwritten rules or hidden labels. This capacity reflects a form of humanity's sixth sense: an intuitive reasoning mechanism that recovers implicit information beyond raw senso

TII has released Falcon-ASR, a 1.6B speech recognition model for Arabic with an Emirati focus. It reports 20.92% average WER on 6 Arabic test sets, but weights and an API are not yet available. The post TII Releases Falcon-ASR: 1.6B Arabic Speech Model Hits 20.92% WER appeared first on MarkTechPost .

Efficient visual tokenization requires balancing representation compactness with the preservation of fine-grained image content. Conventional grid-based tokenizers allocate tokens uniformly across spatial locations despite non-uniform visual complexity, whereas compact 1D tokenizers aggregate global information but can struggle to preserve local details. We introduce MixTok, a hybrid visual tokenizer that combines a compact 1D global representation with sparse 2D tokens for local refinement. A c

What do you use it for, and what maintenance has it needed? I'm interested in services that earn their place over time.

Turn any website into clean data with one API call. Discussion | Link

Hand-drawn diagrams your AI agent can draw with you Discussion | Link

The newly named D-Series GPU has up to 480GB of LPDDR5X memory, 1.5TB/s of memory bandwidth Intel rates it for up to 1.3 PetaFLOPS of FP8 performance, with support for formats ranging from FP4 to FP64. It features a total of 32 Xe3P cores, 256 XMX engines, 32 MB of unified L2 cache. No cost estimates currently, I hope they launch a generic blower style variant at a later date for prosumer workstations. Intel GPU D-Series Specifications Intel GPU D270 **Xe Cores** 32 **XMX Engines** 256 **GPU Clo

I can't find anything in the Docs. I see that the setup.sh file mainly pulls a big ass unaudited binary and, with all simpathy for the repo author, that's not going to run on my machine. I am sure the documentation is somewhere , but I can't locate it. (Yes, I could ask an LLM to search the repo, but from time I want to do things the old way, reading myself, and asking for help to the FOSS community). Where are the docs to build strata from source?