The study proposes FSVL, a fast-slow dual-pathway framework combining a lightweight ResNet50 fast classifier with a YOLOv8n-guided slow path for fine-grained fish identification, integrated with a retrieval-enhanced Qwen3-14B for domain knowledge generation. On 607 test images, the fast classifier achieved 99.67% accuracy, and the cascade corrected errors to 100%; knowledge generation improved scientific name accuracy from 91.37% to 100% and ROUGE-L from 0.335 to 0.655.
A study on Qwen3.5-27B tool-use tasks across 32 fresh worlds finds that case-specific residual-writing interventions lift familiar target-selection correctness from about 27% to roughly 90%, but these gains vanish when familiar facts and identifiers are recombined in the same task, and they yield little improvement in independent eligibility or selection judgments.
Yuval Kansal and coauthors introduce Self-Direction, a mechanism that lets agents reflect on failed trajectories, distill the lesson into a lightweight low-rank adapter of roughly 0.02% of base-model parameters, and store it in a persistent lesson library for automatic routing. On GAIA, Claw-Eval, and Terminal-Bench 2.1 with Qwen-family open-weight models, success rates improve by 8.9 to 17.0 percentage points, repeated mistakes drop, and a handful of adapters generalize to dozens of new tasks.
Hugging Face Bio has released Carbon-X, a 14.8B-parameter decoder-only DNA language model built on Qwen3-14B and licensed under Apache 2.0. It reads raw DNA directly in a chat interface and answers in English, averaging 66.7% across five DNA benchmarks—just 0.8 points behind GPT-6.1 Sol.
Analytics Vidhya hands-on tests Moonshot AI's open-weight Kimi K3.1 Pro, which comes close to Claude Opus 5.5 on agentic coding and raw 3D reasoning, leads on financial document analysis, and sharply cuts hallucinations versus K3. It still trails Opus 5.5 and GPT-6 Astra on tool-assisted 3D work and hallucination control, and full weights arrive November 4, 2026, pending independent verification.
Sebastian Raschka released an 11-hour, eight-video coding course that implements reasoning-model techniques — evaluation, inference-time scaling, reinforcement learning, and distillation — from scratch in Python and PyTorch, using a small pretrained Qwen3 model as the starting point.
Moonshot launched Kimi K3.1 Pro, a multimodal 3T-class mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters, and a 1M-token context window, posting near-frontier independent benchmark results. The API is live across Kimi's website, app, Kimi Work, Kimi Code, and developer API, while open weights are promised for November 4, 2026. Artificial Analysis placed it fourth on its Intelligence Index at 55, three points behind its top three, and said the open-versus-proprietary gap would narrow from 12 points to three once weights ship.
Moonshot AI has launched Kimi K3.1 Pro, the first entry in its K3.1 series: a 2.8T-parameter MoE model with 104B active parameters and a 1M-token context window. It scores 65.7% on Terminal-Bench 4.0, trailing Claude Opus 5.5 by 0.7 points, and is priced at $2/$10 per million tokens, below K3. Open weights are promised for November 2026.
Moonshot released Kimi K3.1 Pro on the Kimi API on 23 October 2026; the weights are not open yet, but the company promises full model weights by 4 November 2026. Pricing is $2 per million input tokens and $10 per million output tokens — matching GPT-6.1 Sol and half the cost of Claude Opus 5.5. Artificial Analysis scored it 55, behind only Claude Opus 5.5 at 58, the highest a soon-to-be open weight model has reached against the frontier.
A new arXiv paper introduces FusionEval, a benchmark for systematically evaluating model fusion techniques for large language models. It builds controlled pools of domain experts trained from Qwen3 and Llama base models across math, code, science, instruction following, and tool use, then measures both capability retention and computational cost on eight public benchmarks, finding that training-free fusion preserves reasoning but substantially degrades tool use.
AWS announced the general availability of training sessions within Amazon SageMaker AI serverless model customization. Users write and coordinate their own post-training loop, including custom loss functions, reward signals, and multi-turn reinforcement learning, while SageMaker AI handles GPU allocation, distributed execution, fault recovery, and checkpoint storage. The feature supports gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and Qwen3.8-27B, and billing is based on tokens processed.
Liquid AI has introduced RIPL, an inference engine built to run its LFM models inside live request paths with end-to-end generation under 50 ms. The company's own benchmarks show 5.6x to 28.4x lower p95 latency than Qwen3.5 on vLLM, and Shopify is already running LFMs at sub-20 ms in production search. RIPL is enterprise-only, closed, and all figures are self-reported.
A new arXiv paper investigates how field labels in structured prompts and auxiliary hints affect visual binding in visual question answering. By holding images and claims fixed across ten models and six tasks, the authors find that field-label effects are model-dependent: adding object-role labels helps Qwen3 but hurts Qwen2 on Spatial. Preserving hints stay near a no-hint baseline, clearer wording improves text matching but not consistently visual judgments, and direct relation selection partly reduces errors for Qwen3 and GPT-4.1-mini.
This paper investigates whether large language models form internal representations corresponding to all 26 CATO factors used in case-based legal reasoning. Experiments on three open models—Qwen3-1.7B, Qwen3-8B, and Gemma-2 2B—examine linear decodability and sparse localization with sparse autoencoders, finding that all 26 factors are linearly decodable with best-layer test AUCs from 0.86 to 1.00, and reporting 78 model-specific layer-feature coordinates.
This paper introduces ReAPO, a simple modification to GRPO derived from viewing credit assignment through the lens of Stochastic Shortest Paths. ReAPO allocates additional generation to failed trajectories so that successful retries are reinforced without rewarding preceding failures, assigning credit to each attempt in a theoretically principled way. Across Qwen3 and OLMo-2 models from 0.6B to 4B and eight mathematical and scientific reasoning benchmarks, ReAPO improves first-attempt performance over GRPO and recent baselines; on Qwen3-0.6B it matches the peak accuracy of GRPO with twice the group size or twice the training steps using 39% and 31% fewer generated tokens, respectively.
Train Qwen3 on AWS Trainium with torchtitan three ways on a single trn2.48xlarge: the Hugging Face Transformers v5 modeling backend, torchtitan's native Qwen3, and NKI mixture-of-experts kernels. The same torchrun launcher and config structure work unchanged across all three paths, including loading and saving published Hugging Face safetensors.
arXiv:2610.27040v1 Announce Type: new Abstract: General-purpose multimodal models, including GPT-6 Astra and Claude Opus 5.5, demonstrate broad capabilities in embodied manipulation. Our analysis indicates that this generality is accompanied by strong spatial perception, as Astra can determine where to interact with task-relevant entities amid visual distractors. Because spatial perception connects high-level reasoning to the physical scene, we investigate scaling visual grounding to strengthen the perceptual foundation of general embodied operation. We develop a unified data pipeline that combines robot manipulation recordings with complementary visual data from diverse sources. After automatically annotating the robot recordings, the pipeline converts the collected data into verifiable…
arXiv:2610.26937v1 Announce Type: new Abstract: Large clinical language models often require multi-GPU infrastructure, limiting local deployment. We introduce CERTA, a decision-level statistical certificate that determines when a quantized model may substitute for its native BF16 reference and escalates decisions that cannot be certified to it. Across eleven instruction-tuned models, both general-purpose and domain-adapted, from three families (7-72B), we evaluated 2,500 clinical-verification decisions from de-identified patient contexts (2,484 patients), with patient-disjoint calibration and testing. In the higher-power operational analysis, round-to-nearest 8-bit quantization at BF16 storage served 99.8-100% of held-out decisions across all eleven models; formal finite-sample guarantees…
arXiv:2610.26866v1 Announce Type: new Abstract: Mechanistic studies of large language models localize function in parameters, attention heads or neurons, but those parts do not specify how one inference unfolds. We treat the ordered layer-state trajectory as the empirical process and test whether layer order defines a measurable computation. Across representative Qwen, Llama and Gemma checkpoints, realized order produced paths 1.60--2.24 times as aligned with their endpoint chord as endpoint-preserving shuffles. The order effect transferred to a lexical task, but was stronger in three random initializations of the same Qwen architecture, identifying an architecture-compatible scaffold rather than a training-specific law. Direct hidden states provided stronger mechanism coordinates than vo…
arXiv:2610.26887v1 Announce Type: new Abstract: Large language model (LLM) agents are evolving from conversational assistants into Cowork Agents capable of autonomous planning, tool use, and complex task execution. Financial analysis requires these agents to integrate cross-document evidence and perform multistep numerical reasoning, but expert task authoring is costly, while direct LLM synthesis struggles to produce challenging tasks with reliable answers. We introduce Navalia, a framework for Cowork Agents that combines verifiable task synthesis with long-horizon post-training. Using corporate financial reports as its data source, Navalia semantically selects supporting evidence from complete fact sets and constructs executable computation chains to generate analytical tasks with tracea…
DeepSeek has open-sourced the inference engine it uses to serve its models in production, a step it said last year The post DeepSeek open-sources the inference engine behind its API. And Huawei gets equal billing with Nvidia. appeared first on The New Stack.
DeepSeek has open-sourced DeepInfer, the Apache 2.0 inference framework that serves DeepSeek-V4.1-Flash in production. It targets NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs, and ships offline inference only. The post DeepSeek Open-Sources DeepInfer, the Inference Framework Behind DeepSeek-V4.1-Flash appeared first on MarkTechPost.
Use of tools ‘may result in the transfer of sensitive data to (China)’, defense department says The US government on Monday ordered its agencies to stop using artificial intelligence (AI) tools from Chinese startup DeepSeek, as the countries race for dominance in the fast-moving sector. A notice from the Department of Homeland Security (DHS) announced the ban on products developed by DeepSeek – and tools created using DeepSeek by third parties – saying they must be removed from IT devices. Continue reading...
Multi-turn agentic workloads reprocess growing conversation context each turn, driving up latency. See how bi-directional KV cache transfer lets prefill and decode nodes reuse previously computed context across turns, cutting time to first token by up to 1.6x for Qwen3-32B and 3.3x for GLM-5-FP8 on Amazon EC2 P5en instances with EFA.
Not content with challenging the major frontier AI labs with lower-cost models, Chinese AI company DeepSeek is setting its sights The post DeepSeek made its name with cheap AI models. Now it’s undercutting search API rivals, too. appeared first on The New Stack.
arXiv:2610.22454v1 Announce Type: new Abstract: Running prompt tokens through fewer routed experts during prefill is a simple way to make mixture-of-experts (MoE) serving cheaper, yet it is unclear what this damages. We locate the damage with token-, layer- and attention-path interventions on short-answer question answering in four MoE families. At half the experts the F1 loss of three families is mostly answer form: the model still finds the answer but stops answering briefly. In Qwen3-30B-A3B this loss is made at the two chat-template tokens before the answer, and in DeepSeek-V2-Lite at the first 8-32 prompt tokens, where its instruction sits; by answer likelihood these positions account for about half and a fifth of the damage. Reduction changes the first token's attention mass, norm a…
arXiv:2610.22381v1 Announce Type: new Abstract: When a fact becomes harder to elicit from a language model, has the model lost it, or has it merely become less accessible? Parameter-updating test-time continual learning agents are usually evaluated by current output accuracy; we ask whether that measurement captures how prior experience shapes later updates. We test whether learning a specific factual binding produces a later relearning advantage beyond familiarity with its entity, answer vocabulary, format, or query slot. In Qwen2.5-1.5B, we compare three matched, counterbalanced histories: LEARNED receives the correct binding, CONTROL receives matched exposure without it, and WRONG receives a competing relation-valid answer in the same slot. After an identical rehearsal-free continual f…
Post-train open LLMs like Qwen, GLM and Kimi into custom models you own, with SFT or RL. You write the training loop. Together runs the hard infrastructure.
In the west the technology comes with at least some effort to provide security and a benign intent – not so from China There is much to fear from artificial intelligence developed in China and handed to businesses and governments in the west almost free of charge. Every month Chinese AI developers such as DeepSeek and Xiaomi dispatch the latest versions of their “open-weight” models that allow users to take the core software and use it without paying an expensive ongoing licence. Continue reading...
Release: web-edit 0.1a2 web-edit (previously) is my tiny web app for browsing and editing files in a folder. I mainly built it for editing files on different machines via Tailscale. I added a "copy to clipboard" option for both text and image data. I needed this because I was running pelican prompts on a laptop from my phone (here's DeepSeek 4.1 Flash 1017) and I wanted an easy way to copy the resulting logs into a Gist. Tags: projects, pelican-riding-a-bicycle, deepseek
DeepSeek has just released DeepSeek-V4.1-Flash-1017, a new checkpoint of its V4.1-Flash model. It is live on the DeepSeek website, app and API. DeepSeek claims stronger scores, faster task completion and shorter outputs than the September V4.1-Flash release. The company calls it a Pareto improvement in performance and cost. Alongside it, DeepSeek opened its Search and […] The post DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API appeared first on MarkTechPost.
arXiv:2610.20627v1 Announce Type: new Abstract: Vision-language models are increasingly used for driving-scene understanding, yet their defensive reasoning under weakly structured Global South traffic remains unclear. We introduce BD-HazardVLM, a diagnostic benchmark of 500 Bangladesh road-scene images, comprising 375 from RSUD20K and 125 from TFP-BD, annotated for hazard presence, primary hazard category, risk timing, severity, defensive actions, visible evidence, and hallucination- and Western-road-bias traps. We evaluate Qwen2.5-VL-7B, InternVL3-8B, and MiniCPM-V-4.6 using a shared zero-shot task definition and structured output schema. Binary hazard-presence accuracy ranges from 88.0 percent to 89.2 percent, essentially matching the 89.0 percent majority-class baseline. Fine-grained r…
arXiv:2610.20633v1 Announce Type: new Abstract: Long-horizon failures of language models are usually measured as end-task success or as a horizon length. We instead score, step by step, three components of a 400-step ledger task in which every step is easy in isolation and the model's own answers are re-injected into its context: recognising a record already seen 5 to 300 steps earlier, processing the current record locally, and writing a counter incremented at every step. On Qwen3-8B (non-thinking, 20 seeds, replicated on a second machine and backend), with every step scored component by component, two results hold. First, duplicate-recognition errors rise late in the episode -- from 0.0019 to 0.011 per step at risk across step 200 -- while the counter makes no error in 8,000 steps and l…
arXiv:2610.20681v1 Announce Type: new Abstract: Formal inequality proving requires mathematical reasoning and exact computation, often connected through intermediate lemmas and goal reformulations. We introduce IneqAgent, an agentic framework combining LLM reasoning with symbolic computation for inequality proving in Lean4. Lemma-level Monte Carlo Tree Search coordinates lemma construction, goal reformulation, and symbolic backend selection. Complementary sum-of-squares, Bernstein, and successive difference-substitution backends produce exact certificates for local Lean proof generation. The agent integrates these proofs with LLM-generated steps and uses Lean feedback to construct complete proofs checked against the original theorem statements. On four benchmarks, IneqAgent with DeepSeek-…
arXiv:2610.19033v1 Announce Type: new Abstract: Low-rank adaptation (LoRA) applies every rank-one component of its update to every token, and mixture-of-experts variants replace this with a learned router that activates a fixed number of components. We argue that the number of components a token needs is a detection question with a classical answer. A token's responses to the rank-one components of a trained LoRA consist of a few large values on a Gaussian noise floor whose width matches that of an untrained adapter, and the floor carries cross-talk rather than adaptation. We introduce SQUELCH, a router-free mixture of rank-one experts that (i) soft-thresholds each token's responses at a multiple of its own robust noise level, so the active rank varies per token; (ii) chooses that multipl…
arXiv:2610.18883v1 Announce Type: new Abstract: Token-selection methods for on-policy distillation (OPD) prioritize student-visited positions using diverse signals, including uncertainty, teacher-student disagreement, and mismatch direction. We instead derive token importance from the distillation objective itself: how much can each position's loss decrease under an equal local KL budget on the student's predictive distribution? The answer is the utility $s_t=g_t^\top F_t^\dagger g_t$, where $g_t$ is the logit gradient and $F_t$ is the categorical Fisher matrix. For reverse-KL it equals the student-weighted variance of the student-teacher log-probability ratio, and for forward-KL it equals the Pearson $\chi^2$ divergence, so the appropriate utility follows from the choice of objective. On…
arXiv:2610.18878v1 Announce Type: new Abstract: Post-training reasoning LLMs often requires making a tradeoff between costly supervision and compute inefficiency: supervised finetuning typically requires reasoning traces, which is data inefficient since in many applications one only has prompt-answer tuples and lacks the groundtruth chain-of-thought (CoT) sequence. On the other hand, reinforcement learning (with verifiable rewards, for example) can operate without CoT data but it is much more compute-inefficient since it requires multiple autoregressive completions per optimization step, none of which are guaranteed to give a non-zero policy gradient. In this work we explore a middleground, a self-distillation technique that uses answer-only data, treating CoT as latent variables. This re…
arXiv:2610.14190v1 Announce Type: new Abstract: Explaining a vision--language model (VLM) prediction requires tracing training-example influence to the multimodal evidence that mediates it, yet existing attribution methods largely operate only at the sample level. Recent latent-mediated influence methods provide feature-level attribution in language models, but lack the multimodal and spatial structure needed to ground influence in VLMs. We introduce MOSAIC (Multimodal Orthogonalized Sparse Autoencoders for Influence Characterization). It learns a shared sparse image--text latent space whose concepts are 1) sparse within each modality, 2) supported on contiguous image regions, 3) aligned to the phrases that describe them, and 4) approximately decorrelated. It then applies EK-FAC to decomp…
arXiv:2610.14118v1 Announce Type: new Abstract: The rapid advancement of Large Language Models (LLMs) has driven their deployment toward general-purpose harness environments, where LLM-based agents operate in sandbox workspaces and interact directly with file systems and shells. While numerous post-training pipelines exist for tool-use agents, their counterparts for harness-based agents remain scarce and are predominantly one-shot and teacher-dependent: a stronger external model synthesizes tasks and collects trajectories, the target agent is trained once, and evaluation results are consumed by human researchers rather than fed back into the loop. Such pipelines are inherently bounded by the capability of the external model and the scope of human prior, leaving no room for self-improvemen…
The social media giant tries to capitalize on, monetize its successful personal agent and maybe crack the business market. Chinese models go astray, like their U.S. counterparts.
A new case of an out-of-control AI model shows that enterprises should treat models and agents as toddlers that need supervision and explicit instructions.
Release: llm-gemini 0.35 I added support to llm-gemini for the new Gemini 4 Argon model, made available today via the Gemini API and for customers of their $100/month subscription plan. Here are the low, medium and high pelicans. As expected, they are all excellent: Gemini has a track record of producing great SVGs, and are the only lab where I know their staff have invested in SVG specifically (and who have boasted about their pelicans before.) Here's the "high" illustration, which cost me 17.166 cents: The pricing is interesting. At launch, Argon is $2/million input and $10/million output - the same price as GPT-6 Sol, half the price of Claude Opus 5.5 and 1/5th the price of Claude Fable 5.1 and GPT-6 Astra. But on December 1st Argon is set to double in price - so it will be the same pr…
Alibaba’s Qwen team has introduced QwenGyre, an elastic reinforcement learning RL framework for extreme-long-horizon agent training. QwenGyre is an online reinforcement learning system from Alibaba that trains coding agents through unmodified, black-box harnesses. It targets rollouts that run for hours and process close to 1M tokens. The QwenGyre elastic RL framework was used on Qwen’s […] The post Alibaba Qwen Introduces QwenGyre: Elastic Reinforcement Learning RL for XLong-Horizon Agents appeared first on MarkTechPost.
arXiv:2610.12602v1 Announce Type: new Abstract: Continual learning (CL) seeks to adapt pretrained models while preserving prior capabilities. Yet the tasks supplied for preservation cover only part of what a model knows, while forgetting can extend beyond them. Recent work shows that language models can supplement this limited evidence by generating useful replay without access to their original training data. This raises a question about self-preservation: can a model help choose preservation targets that protect its own capabilities more broadly? To study this question, we organize task descriptions and their examples into Task Banks, whose entries can be selected and combined to guide replay generation by a frozen model. Continual instruction tuning with Qwen models of varying sizes yi…