Daily report
Only content delivered through the publication boundary on this date is included.
Sealed edition v1 · 1 withdrawn or replaced articles omitted
BLIS is a deterministic, CPU-only discrete-event simulator that models the whole LLM inference serving pipeline, including admission, routing, scheduling, prefill/decode execution, KV-cache management, and preemption, reaching 6.7% median end-to-end error across 36 experiments while running about 200 times faster than real execution.
Anthropic let Claude design quantum circuits in a closed loop with no human input between rounds: given plain-language specifications and an automated evaluator, the model produced new designs for eight computational primitives spanning chemistry, optimization, physics simulation and number theory, using both fewer qubits and fewer gates than the best previously published circuits, in some cases by an order of magnitude or more. On the elliptic-curve point-addition task it worked for 53 hours with minimal human input and produced the first reversible divsteps implementation with no stored branch history; its Reed–Solomon decoder for DQI cut gates 12-fold relative to Google's previous construction.
Self-improvement in scientific agent swarms no longer rests on one agent's repeated trial and error. Previously, a failed experiment could not show whether the hypothesis or its implementation was at fault, and a program without overall gains could still hide a component another agent needed. SwarmRSI places hypotheses, programs and experimental evidence into a shared graph as inheritable research states, uses other agents' feedback to decide whether to revise a hypothesis or repair a program, retests retained hypotheses on other agents' programs, and combines complementary components. It reaches an average of 76.2 on OpenProblems and matches published SOTA on 44.4% of NatureBench tasks, the highest rate among compared methods, with medals on 81.8% of MLE-Bench Lite tasks.
The provenance of the NIST standard elliptic-curve “seeds” had been unexplained since the NSA supplied them without justification, and the community long suspected possible backdoors. An internal Anthropic research model, Claude, found in under half an hour that the seeds for P-192, P-224, P-256, P-384 and P-521 are not the output of any hash algorithm but of the 1990s C library generator lrand48 seeded with a Unix timestamp, with the P-256 seed appearing at iteration 1842. This is not an attack or a weakness in ECDSA; it makes deliberate backdoors much less plausible, while the B-163 seed and the base points of P-224, P-256, P-384 and P-521 remain unexplained.
The limits of token-by-token generation are now fully characterized: a sampler with little private state cannot sample a pseudorandom generator's output, while an insertional sampler that may write at any position can.
Echo turns the tensor-parallel all-reduce into 4-bit transport with a high-precision sum and feeds each rounding error back into later calls so the accumulated error stays bounded with depth, running 2.09x faster than the 16-bit NCCL all-reduce on 1 GB messages while Qwen3-8B stays within 0.4 points of 16-bit inference across seven benchmarks.
PPO's clipping in LLM reinforcement learning is generalized into a two-dimensional design space of divergence geometry and estimator topology, where KL divergence with atomic estimation raises average scores by about 50% on math and 14% on coding over standard GRPO across eight competition-level benchmarks.
NVIDIA released Nemotron 3.5 Super VL, an open, customizable multimodal model for long enterprise agent workflows that need reasoning, tool use, and text, image, document, and video understanding. It uses a 120B-parameter mixture-of-experts architecture that activates 12B parameters per token, and NVIDIA reports up to 5x higher throughput on agentic tasks than other open models in its size class on a single GPU. Teams can post-train it with NeMo on their own data, tools, and workflows, and NVIDIA also released agentic tool-use, finance reasoning, and 40 vision-language datasets.
A study of text, vision and audio mixture-of-experts pretraining fits a data-mixing scaling law from 268 runs and draws actionable conclusions: audio should spend compute on parameters, repeating vision or audio data beyond two passes buys little, and vision interferes most with audio.
Across 18 Base–Instruct model pairs, harmfulness is more decodable from attention updates while refusal evidence is written mainly by MLPs, and post-training strengthens that writing without improving recognition.
Zebra3D swaps dense global attention in pretrained 3D reconstruction models for a per-layer mix of linear and tile-sparse attention, holding pose accuracy close to dense models at roughly 94% mean attention sparsity and reaching up to 5.6x end-to-end speedup at 1,024 views.
A multi-bit LLM watermark's account identifier can be displaced without the secret key: an attacker with a few accounts of its own can make the decoder read a text as belonging to a different account.
Gamma replaces pixel keyframe memory with auditable text lines, raising average success on RoboMME's sixteen memory tasks from 44.5% to 66.0% under a fixed execution policy and matching the privileged subgoal ceiling on three tasks.
Sustained concurrency in asynchronous RLVR rollout concentrates context processing into fewer decode steps, making the KV cache rather than weights the dominant cost. KV4-RL stores the full context in calibration-free INT4 and protects only attention sinks and recent context in bounded INT8; across full-parameter GRPO training on Qwen3-4B-Base and Qwen3-8B-Base, its six-benchmark mean accuracy stays within 0.5 percentage points of BF16 rollout while substantially accelerating rollout and full iterations.
In multi-agent coordination, deciding whether a coordinator may act on its inferences by inference accuracy fails in both directions, approving harmful language-model deployments in Hanabi where shared errors compound and rejecting beneficial ones in task assignment where costly errors are rare.
DiTaR builds video rewards on task-relevant latent dynamics, reaching 94.3% final success on MetaWorld-v3 versus 65.2% for the strongest video-reward baseline ReWiND, and raising real-robot offline RL success from 43.3% to 83.3%.
Looped language models used to trail a non-recurrent model of the same size at one loop; a single checkpoint now matches a non-recurrent baseline trained on the same data at one loop and gains 7.25 points on a ten-task average at four loops.
Coding-agent evaluation becomes a duel in which two agents author and solve each other's tasks: over 1,500 turns, GPT-5.5 with Codex CLI repaired 98.3% of the defects it reviewed, while Claude Opus 4.8 with mini-swe-agent was the strongest task author with a 27.3% Red win rate.
Robustness tests for tool-using agents can now check, before the agent runs, whether an edited test case still tests the same task; rescoring the same 14,336 stored trajectories under executable task contracts drops Mistral's attacked-task count from 64 to 26 under a MiniLM cosine threshold.
MemForm splits how an LLM agent's long-term memory is formed into five intervenable interfaces, and across six models and two response formats all five showed behavior being rewritten.
Offline demonstrations, current-policy rollouts, and historical rollouts can now be trained under one regression target, and on four math-reasoning benchmarks the unified configuration improves Qwen3-1.7B benchmark-averaged avg@32 and pass@32 over two-stage SFT plus DAPO by 2.33 and 4.44 percentage points.
Personal agents execute proposals users never accepted, and a lightweight memory add-on that leaves the model and memory system untouched cuts that unauthorized execution from 0.44 to 0.08.
Evo2 turns harness edits and their execution outcomes into training signal for the model that writes them, lifting SWE-bench Verified resolution from the strongest baseline's 41.0% to 48.3% and Polyglot pass@1 from 62.6% to 82.1%.
Reasoning models draft code early inside their chain of thought, and SecReason scans the streaming tokens before delivery and inserts a warning so the model repairs the draft within the same reasoning trace, raising the share of runs that are both secure and functionally correct from below 15% to 72.8%–98.1%.
Recording one augmentation strength per instance and letting content variances respond to it lets self-supervised learning separate content coordinates that previously stayed mixed, raising dSprites factor alignment from 0.837 to 0.932.
On the public AgentX agent workload, an eight-GPU AMD Instinct MI355X system serving Kimi K3 with vLLM reaches about 7,567 total tokens/s/GPU at 50 tokens/s/user P90 interactivity, 24.4% above NVIDIA HGX B300's 6,083; under SemiAnalysis's modeled $1.50 versus $2.26 per GPU-hour, that is 18.16 million versus 9.69 million tokens per dollar of TCO, 87.4% more. The result comes from sustained AMD and vLLM community engineering across attention, recurrent state, MoE and caching rather than a single change.
NVIDIA released three open Nemotron 3.5 models: the 120-billion-parameter Super VL for multimodal agentic tasks, which reaches leading accuracy on instruction following, tool use, multistep business workflows and visual understanding comparable with NVIDIA's largest reasoning model at roughly one-fifth the parameter count, plus the edge-oriented Bolt 74B-A9B 24GB and Bolt E4B 4GB. Super VL and Bolt 74B-A9B are available now, with Bolt E4B coming Oct. 23, covering deployments from enterprise servers down to local hardware.
Reflection's Beam is a 501B-parameter mixture-of-experts model with 23B active parameters per token, and vLLM supports it from day one with reasoning, tool calling and structured outputs working out of the box. Its first implementation was almost pure PyTorch; coding agents profiled the model, wrote kernels and benchmarked each change, cutting batch-1 decode on TP8 from 8.95 to 3.30 milliseconds per step and an 8×2K prefill from 328 to 176 milliseconds. With its DFlash2 draft head, a single request reaches 708 tokens per second on TP8 at an acceptance length of 3.7.
Malignant cell states coexist within a tumor, and single-cell RNA sequencing can resolve them but is costly and hard to scale, while bulk RNA sequencing is cheap but mixes their signals together. CDState uses nonnegative matrix factorization with a sum-to-one constraint on proportions and a cosine-similarity penalty to infer malignant cell states from bulk RNA without reference data; on simulated bulk data from five cancer types it outperformed existing unsupervised methods overall in estimating state expression and proportions. Applied to 9,669 samples across 33 TCGA cancer types, it yielded 193 malignant cell states, with epithelial–mesenchymal transition, MYC targets, and oxidative phosphorylation the most recurrent programs.
Generative enzyme design has mostly optimized at the amino acid level, rarely accounting for coding sequences and host expression at once. EZPro-Fit places protein 3D structure, coding sequences and taxonomic information in one generative framework, pretrained on thermophilic microbial data and fine-tuned on highly expressed E. coli genes, so it generates host-adapted coding sequences directly from structure. On the industrial endoglucanase Cel5E, its variants N291D and D259E raised both melting temperature and catalytic activity in experiments.
The difficulty of agent evaluation now comes from the environment rather than from swapping tasks. Recursive Environment Scaling (RES) rewrites information access, resource relations and execution dependencies while preserving each task's objective and completion requirements, validating every candidate by execution and letting accepted changes be inherited across rounds. On 294 initial tasks from three benchmarks, task success fell for all six evaluation models, by 57.14 to 77.14 percentage points on Claw-Eval-Live; with equal trajectory counts and identical optimization settings, fine-tuning on RES trajectories beat original-trajectory SFT in all nine model-benchmark comparisons by 7.00 to 14.61 percentage points.
Reflection released the weights of Beam, its first open-weight model, on Hugging Face under an Apache 2.0 license; Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning, and long-horizon agentic workloads. Previously the model was reachable only through Reflection and its partners, and developers can now download and self-host it, evaluate it, and customize it for their own domains. The release ships with final benchmarks in the model card and a safety technical report, plus an ecosystem of over 40 partners and integrations spanning inference, quantization, fine-tuning, reinforcement learning, agent harnesses, evaluation, and observability, which Reflection calls an open operating system for agents.
A UAV ISAC synthetic aperture radar imager casts imaging as an exact linear inverse problem, using a pretrained diffusion model as prior and the echoes as likelihood guidance to reconstruct super-resolution images under arbitrary known trajectories and unknown position errors.
Converting plane-wave DFT Hamiltonians into atomic-orbital form now uses projectability-selected eigenstates to complete the subspace instead of an empirical shift, giving 1–3 meV band errors for five sodium phases with 13 orbitals per atom where real-space reconstruction needs 23 orbitals and still errs by 12–28 meV.
Plasma cell-free DNA methylation sequencing, long used only to read methylation, is shown to partially retain fragmentomic signals such as fragment length and end motifs. The study delineates co-methylated blocks directly in plasma cfDNA, places methylation and fragmentomic features in the same genomic intervals, and builds an interpretable ensemble model for cancer detection and tissue-of-origin prediction. Cross-platform comparisons show major fragmentomic patterns are recapitulated in methylation sequencing, though preservation varies by feature.
Decisions in colorectal liver metastases weigh tumor biology, resectability and future liver remnant, and large language models previously produced recommendations from prompting alone with limited agreement with expert multidisciplinary practice. The METATRON study formalized the evaluation sequence hepatobiliary surgeons follow at tumor boards into a structured workflow and combined it with guideline retrieval, testing 110 synthetic vignettes scored by three independent LLM judges across five dimensions. The workflow-plus-retrieval strategy reached the highest composite score (0.626 versus 0.444 for baseline), yet sensitivity for anatomical resectability remained 40%, insufficient for autonomous clinical use.
Vessels reconstructed from low-resolution contrast-enhanced CT can come out broken, spurious, or wrongly connected, and image-quality metrics do not capture those structural errors. VascuSR separates vascular structure from CT intensity reconstruction: a conditional 3D rectified-flow refiner corrects vessels in label space, a lightweight controller injects the refined structure into the diffusion decoder, and each new super-resolved volume re-extracts the structure for the next round. On re-annotated MSD Task08 and HVM, it beats interpolation and learning-based baselines in Dice and clDice, cuts class-averaged connected-component counts by over 20% on both datasets and mean centerline distance by over 10% on MSD, while keeping image quality comparable to its SR backbone.
Reflection released Beam, its first open-weight model, with 501 billion total parameters and 23 billion active in a sparse Mixture-of-Experts design for coding, reasoning and long-horizon agentic workloads, publishing weights on Hugging Face under Apache 2.0 with a model card and a safety technical report. Where an open-weight release usually delivers one model layer, this one also ships BF16, FP8 and NVFP4 precision formats, a speculative decoding drafter, and integrations from more than 40 partners across inference, fine-tuning, deployment and agent harnesses, so developers can download, self-host and customize it for their own domains.
Kimi K3's decode path in vLLM is now a unified CuTe DSL implementation built with agent assistance, fusing attention, MoE and cross-GPU communication. On 16 NVIDIA GB300 GPUs with DSpark at synthetic acceptance length 6, decode reaches 842 tokens/s, 1.73x stock vLLM under the same speculative setup, and without speculation TP16 keeps 1.62-1.71x at batches 16-64. A proposed fast-track process pairs such work with isolated changes, bounded hardware scope and both kernel and model-level validation.
Quantum DeepONet ensembles cut the hidden-layer cost of operator learning from the quadratic cost of dense layers to near-linear, and add a finite-sample, distribution-free lower bound on query-averaged coverage. Previously quantum operator learning returned only noisy point predictions, while Bayesian networks and Monte Carlo dropout offered no distribution-free guarantee. The approach replaces dense hidden layers with quantum orthogonal neural networks and applies split conformal calibration to the mean and standard deviation of independently trained ensembles. On synthetic antiderivative and advection benchmarks, relative L2 errors are 0.46% and 2.28%, with 8-model ensemble coverage of 92.13% and 88.99% against a 90% target.
At the Dresden University Cancer Center's MASTERsg precision oncology program, a gradient-boosting model re-reads the raw bin-level copy-number data from TSO500 panel sequencing to find focal PTEN deletions that the standard pipeline misses. TSO500 averages read counts across bins within a gene, so focal events fall short of its fold-change threshold; the model flagged deletions in 33 of 204 samples classified as PTEN wild-type, and 15 of the 20 stainable cases showed absent or reduced PTEN protein.
Meta released Muse Spark Open, an open mixture-of-experts model for agentic coding with 816B total parameters, 42B active per token, and a 1M token context window that holds an entire project including complex dependencies and legacy code. General coding models are optimized for single-shot completions such as generating a function or a unit test, where context is small and the answer immediate, while multifile refactors, navigation across thousands of files, recovery from failing tests, and subagent coordination require sustained multistep work. Trained on NVIDIA GB300 NVL72, the model supports tool use, output evaluation, and iterative refinement, using goal conditioning to hold direction and context compaction to sustain progress across long sessions.
NVIDIA released Nemotron 3.5 Super VL, an open multimodal model for long-running agents that folds document, image, and video understanding together with reasoning, instruction following, and tool use in one model, reducing the need to deploy separate text and vision models. It uses a 120-billion-parameter mixture-of-experts design with 12 billion active parameters, combined with a hybrid Mamba-Transformer architecture, multi-token prediction, and NVFP4 quantization, and can be post-trained with NVIDIA NeMo on an organization's own data, tools, policies, and workflows. The stated comparison is up to 5x faster throughput on agentic tasks than other open models in its size category on a single GPU, and 4x faster task completion.
Cold air outbreaks over East Asia are now sorted by which side of Mt. Baitoushan the cold airmass flux passes, not by intensity alone. Using JRA-3Q reanalysis for winters from 1979/80 to 2023/24, researchers computed the flux of cold airmass and applied self-organizing maps to group 2,052 outbreak cases into Western, Northern, and Intermediate groups, then defined a pathway index as the log ratio of flux west and east of the mountain. The index correlates negatively with 925 hPa wind convergence over the Sea of Japan, so the western route favors the Japan-Sea Polar airmass Convergence Zone and heavier coastal snowfall.
Anthropic's Claude Platform release notes say fast mode is out of research preview for Claude Opus 5.5: developers no longer request access, requests no longer need the fast-mode-2026-02-01 beta header, and requests that still send it keep working. Claude Opus 5.5 now carries a default fast mode rate limit in the Start, Build, and Scale usage tiers, and fast mode reaches Claude in Microsoft Foundry deployments hosted on Azure. Fast mode for Claude Opus 5 and Claude Opus 4.8 remains in research preview and still requires the beta header.
Reflection released Beam, its first open-weight model, with day-0 support on the Baseten platform. Beam is an autoregressive transformer with a mixture-of-experts architecture: 501B total parameters, 23B active per token, and a context window up to 1 million tokens, aimed at coding and agentic workloads. Training leaned on reinforcement learning, using longer multi-step rollouts so the model can practice tool use and adjust from feedback, plus a length penalty against unnecessary tokens, with the goal of cutting the cost of these workloads.
Digital restoration of Dunhuang murals no longer denoises only in pixel space. Pixel-domain diffusion must handle global structure and local brushwork at once, which is costly and can drift in color and style. WDCRF first applies a 2D discrete wavelet transform to split a damaged mural into one low-frequency and three high-frequency subbands, then lets a conditional diffusion model rebuild all four coefficient sets directly in the wavelet domain, using the damaged coefficients themselves as the conditioning signal, and finally synthesizes the image by inverse wavelet transform. On the Dunhuang Mural dataset it reaches 32.04 dB PSNR and 0.922 SSIM, ahead of EdgeConnect, WaveDiff, PGRDiff and LRDiff; on DhMurals it reaches 18.95 dB PSNR and 0.7598 SSIM.
Reflection released Beam, its first model, an open weights system with 501B total and 23B active parameters that scores 30 on the Artificial Analysis Intelligence Index, second among US open weights models behind Meta's Muse Spark Open (39) and ahead of NVIDIA's Nemotron 3.5 Super VL (26). It uses about 26k output tokens per Intelligence Index task, roughly half of K2 Horizon (31), while proprietary models still set the frontier, with GPT-6 Astra (max) scoring 53 on the same budget.
IBM released Granite Switch 4.1, which composes general decision-making, safety decision-making and speech transcription into a single hosted endpoint, with 8B and 30B versions occupying roughly the same GPU VRAM as the base model alone. The earlier Granite Switch project already allowed users to compose specialized adapters into custom models, while this release ships a Decision adapter, a Guardian safety adapter and Granite-Speech-5.0-Turbo-CTC precompiled into the model. Because the adapters are built as cache-efficient Activated LoRAs, the KV cache produced by a base-model generation can be reused for later adapter-based decisions, so switching between chat, decision and spoken requests does not rebuild the cache.
Ozone concentration forecasting now runs on two-stage decomposition and two-scale feature selection. The raw series is first decomposed by ICEEMDAN, the most complex component IMF1 is decomposed again by VMD, and the remaining subsequences are reconstructed by fuzzy entropy; partial autocorrelation and random forest then select input dimensions before a temporal convolutional network predicts. On daily ozone data from Beijing and Shanghai, the combined model reaches MAE, RMSE and SMAPE of 0.53509, 0.73346 and 0.8433%, the best among four comparison experiments.
ElevenCreative launched Instant Soundtracks, which generates an original score for an uploaded video, then analyzes the footage for sound-effect cues and places each generated effect at the matching timestamp on the timeline. Previously music and effects were produced separately and aligned by hand; this update also tightens timing for short-form content and mood switching across scenes, and lets the video's existing audio shape the new score through its genre, crescendos, beats per minute or instrumentation. Effects appear as separate layers in the Music Video Editor, and the API adds add_sfx, separate_tracks and use_audio_as_reference, supporting up to 10 videos per request, 200 MB and 10 minutes combined.
At Ship 2026, Vercel shifted its platform from frontend hosting toward agent infrastructure, announcing that Python is now generally available on Vercel and introducing Vercel Daemons in public beta for continuously running background workers. The platform also pairs agent-facing command-line access, connectors, and persistent workspaces so agents can deploy software, be deployed, and take part in operating the platform itself. The stated scope covers running FastAPI, Flask, or Django with standard CPython, with the platform keeping a set instance count and replacing failed workers.
Tessl's engineering team runs a software factory on one monorepo, producing over 1,200 agent-written pull requests a week, and review had become its largest variable cost. The team encoded its own coding conventions, design system and AI code review policy into Tessl Code Review and set the review policy itself, so about 75% of PRs merged without a human correcting the agent or discussing the engineering. In a one-week side-by-side run over more than 600 PRs, 91% of its findings were real defects against 68% for the previous AI reviewer, with 1.56 versus 0.75 author-fixed bugs per PR and about 35% lower review cost per PR.
Perplexity's Decider V1.2 27B and Decider Flash V1.2 4B, two multimodal decision models, are now available through AI Gateway. They classify content, route requests and score text, structured records and images by answering typed questions with probabilities: boolean questions return the likelihood of yes, choice questions pick from categories the caller defines, and score questions assess content against an ordered rubric. An application can use those results to route a support ticket, flag a damaged product, or send an uncertain decision to human review, and several questions can share one input in a single request.
Google has released a repository called Android skills that packages AI-optimized, modular instructions and resources as open-standard agent skills, markdown SKILL.md files that give agents a technical specification of a task plus information on specialized domains and workflows. Development focuses on use cases and workflows where evaluations show LLMs underperform, rather than well-established areas such as basic Jetpack Compose best practices. Skills install through the Android CLI into the current directory or all detected agent directories, or through the Claude Code and Codex plugin marketplaces, and agents activate relevant skills automatically.