Wednesday, July 8, 2026

LLM agents solving multi-step tasks often commit to doomed trajectories while wasting inference compute. This paper proposes an early abort cascade using lightweight probes on hidden activations to predict failure as early as the first round, with recall-controlled gates ensuring successful episodes pass at a user-specified global rate. On TextCraft, at 90% recall target, it saves 47.1% (Qwen-2.5-7B) and 37.2% (Llama-3.2-3B) compute, outperforming single-gate and behavior-only policies by 1.6-1.7x.

PACR-Video is a parameter-efficient framework for multi-shot long video extrapolation, keeping the text-to-video diffusion transformer frozen and augmenting it with low-rank temporal adapters and prompt routing. It builds a recursive prompt bank storing entity, location, action, and style prompts, then routes them via adapter gates according to narrative dependencies. Across six multi-shot and long-video benchmarks, PACR-Video outperforms baselines on distributional quality, semantic alignment, identity consistency, temporal smoothness, motion stability, transition coherence, and human preference, demonstrating that compact prompt routing and lightweight temporal adaptation provide sufficient controllable capacity for stable long video generation.

On July 7, 2026, Meta launched Muse, a new AI image generation model applicable to advertising creative, interior decoration, and creator tools. Trained on large-scale image-text data, Muse supports high-fidelity generation and style control, aiming to provide businesses and creators with more efficient visual content production. This move further solidifies Meta's position in generative AI and may significantly impact advertising and social media content ecosystems.

This paper reveals the root cause of modality interference in full-duplex Spoken Language Models (SLMs) through fine-grained analysis: gradient conflicts arising from shared deep parameter spaces between acoustic and semantic modeling. It proposes Lychee-FD, a framework with hierarchical parameter separation to decouple conflicting modalities and a semantic alignment channel for cross-modality coherence. Experiments show +7.4% on Spoken QA and +28.5% on FullDuplexBench 1.5 without compromising inference efficiency. This is the first work to uncover and solve the modality interference root cause for full-duplex SLMs.

Recent LLM-based TTS systems achieve remarkable naturalness but rely on implicit end-to-end generation, resulting in coarse-grained control. In scenarios demanding precise stylistic interventions and strict temporal alignment, such as audiobook narration and video dubbing, the inability to explicitly manipulate word-level acoustic attributes remains a critical bottleneck. This paper proposes WordVoice, a unified framework: first, it constructs WordVoice-5A, a 4.7k-hour bilingual dataset with five-dimensional word-level annotations (duration, boundary, energy, pitch, and tone) developed through a linguistically-guided pipeline; second, it introduces a bound-token mechanism to transform implicit generation into an explicit controllable paradigm, enabling adaptive multi-task prosodic planning and flexible manual intervention; third, it augments the token-to-waveform stage with a fine-grained acoustic modulation module to bridge the resolution gap between highly compressed discrete tokens and continuous waveforms. Experiments show WordVoice achieves superior decoupled control over multiple acoustic dimensions while maintaining competitive zero-shot synthesis stability. Code and audio samples are publicly available.

This paper rigorously proves a learning separation for quantum machine learning over classical algorithms in predicting time-evolution of quantum many-body systems from a PAC-learning perspective. The authors devise a supervised learning problem with randomized stabilizer probe states, uniform time samples, and observable expectation values, and provide an efficient quantum procedure that learns the underlying Hamiltonian from short-time training samples and performs inference via Hamiltonian simulation and classical shadows. By embedding a BQP-complete computation into a low-intersection variant of the Feynman-Kitaev clock Hamiltonian, they show that no randomized classical polynomial-time algorithm can fulfill the learning condition unless BQP⊆P/poly. The results forge connections between quantum learning theory, quantum simulation, and quantum machine learning, with implications for learning-assisted certified quantum simulation.

Mistral AI announces Robostral Navigate, an 8B-parameter navigation model achieving 76.6% on R2R-CE using only a single RGB camera, without depth sensors, LiDAR, or multiple cameras. By minimizing sensor requirements, it reduces robot navigation costs and deployment barriers, potentially advancing embodied AI practicality.

Anthropic released Claude Code v2.1.203, introducing a login expiry warning, a manual permission mode badge, and MCP roots support, while fixing critical issues such as 15-20s stalls on macOS due to false low-memory detection, permanent unresponsiveness of background sessions after token expiry, lost subagent work, memory/CPU regressions per turn, and improving responsiveness and binary size.

Claude Code v2.1.204 fixes a bug where SessionStart hook events were not streaming in headless sessions, causing remote workers to be idle-reaped mid-hook. This update improves reliability for automated CI/CD and long-running tasks. All users of headless mode are recommended to upgrade.

Wayflow is an embeddable visual workflow editor for the web. It provides AI-powered assistance when desired, and falls back to plain logic when not — compatible with any framework. Open source.

Simon Willison released an experimental Web Component called github-code, generated by GPT-5.5, that lets you embed a specific range of lines from a GitHub repository via a simple HTML tag. The component automatically converts standard GitHub URLs to raw.githubusercontent.com links, fetches the code using fetch(), and displays line numbers. It currently lacks syntax highlighting, but offers a clean way to reference code in blogs and documentation.

Anthropic's Claude Cowork now supports mobile and web, allowing users to start tasks on desktop, check status updates on phone, and retrieve finished output later even with laptop closed. This update makes AI coding agents more seamless for office collaboration, signaling that the 'coding agent wars' are spilling into broader office scenarios.

CLRK (Cognitive Loop Runtime for Kubernetes) is an open-source AI Agent runtime that leverages gVisor sandboxing and MitM (Man-in-the-Middle) guardrails to provide secure isolation and trusted execution for agents in Kubernetes environments. Open-sourced on GitHub, it supports cognitive loop architecture, suitable for high-security automated agent deployments.

Discord admitted that a bug in its AI moderation system had wrongfully banned a large number of users since May, with an additional 200 users banned over the past weekend before the team identified and fixed the issue. The incident has raised concerns about the accuracy and transparency of AI moderation.

sqlite-utils 4.0, the 124th release and first major version bump since 3.0, introduces database schema migrations, nested transactions via db.atomic(), and compound foreign key support. Migrations are defined in Python files using the sqlite-utils library, with a powerful table.transform() method. The release also includes breaking changes and automatic column type detection for CSV/TSV imports.

AI chip startup SambaNova has raised $1 billion in the first close of its Series F round at an $11 billion valuation, just five months after its last mega round. This follows rumors that Intel attempted to acquire the company for about $1.6 billion. The round underscores surging demand for high-performance AI chips and SambaNova's strong appeal in capital markets.

Microsoft announced it will rely more on its own AI models to reduce dependence on third-party models and lower operational costs. This move marks a shift for Silicon Valley giants from an AI arms race to cost optimization, potentially prompting industry-wide imitation. Having invested heavily in OpenAI, Microsoft now pivots to in-house models for greater efficiency.

Former OpenAI executive Kevin Weil has joined the board of Stoke Space, a reusable rocket startup, signaling Silicon Valley's growing interest in space technology. Weil, who led product strategy at OpenAI, is now crossing into aerospace, suggesting reusable rockets are becoming a new magnet for talent and investment. Stoke Space focuses on developing orbital-class reusable launch vehicles and recently completed key technology tests.

Despite the rapid proliferation of open-source models like Llama, frontier labs such as Anthropic have not lost market share. The article argues that the two are actually capturing different phases of the same lifecycle: open-source models accelerate diffusion and application deployment, while frontier labs focus on building stronger foundational capabilities and safety moats. This complementary, rather than competitive, dynamic allows Anthropic to sustain growth through differentiated advantages such as Claude's complex reasoning and safety alignment.

This paper provides a spectral analysis of attention-based graph denoising, showing that linear attention can only learn an average spectral filter, creating a limitation due to spectral diversity across the distribution. The authors propose Spectral Attention and its practical realization, Graph Convolutional Attention (GCA), which performs spectral denoising via graph-filtered queries and keys. Theoretically, GCA matches idealized Spectral Attention under stochastic block models, and the softmax operation is shown to further denoise by projecting noisy eigenvectors onto the clean eigenspace. Empirically, GCA consistently improves graph denoising and diffusion on synthetic and real datasets, and when used with DiGress, it matches standard graph-transformer performance without expensive structural features.

Current benchmarks for evaluating LLMs in data analysis often fail to reflect real-world settings, focusing on fact retrieval from small tables while ignoring multi-table datasets, external knowledge integration, and exploratory insight discovery. This paper introduces DataGovBench, a benchmark from governmental open data, with Table QA and Table Insight tasks. Experiments show significant performance gaps in state-of-the-art LLMs, even with agentic frameworks, indicating they are far from meeting real-world data analytics demands. Code and data are publicly available.

GraphBU is a novel graph-native generator for Mixed-Integer Linear Programming (MILP) instances, where the basic unit is a local subproblem plus its interface to the rest of the instance. By promoting coupling nodes into master constraints or boundary variables and performing compatibility-checked block replacements, it preserves the key graph structure and feasibility of source families. Experiments show that on multiple datasets, the average graph-statistical similarity reaches 0.934, average feasibility is about 96.7%, and the downstream Predict-and-Search training improves its main index by 8.0% on average. This addresses the challenge of obtaining private MILP instances while maintaining structural fidelity.

Long-context language model inference is increasingly limited by the memory bandwidth and capacity required to store key-value caches, yet existing compression methods often apply uniform budgets across layers or tokens and degrade retrieval when lexical cues and semantic states require different preservation. We introduce DepthWeave-KV, a token-adaptive cache compression method that factorizes key and value states across neighboring transformer layers using shared low-rank channel bases while retaining lightweight token-specific residuals where attention behavior is sensitive. DepthWeave-KV combines cross-depth residual factorization with a token-conditional depth router that allocates higher reconstruction rank to instruction-bearing and retrieval-critical tokens, and uses calibration-free online error tracking from attention-output probes to adapt compression during generation without retraining the base model. A fused CUDA implementation jointly performs basis lookup, residual dequantization, and attention projection to reduce decode-time memory traffic. Across LongBench, Needle-in-a-Haystack, L-Eval, and long-form QA and summarization benchmarks, DepthWeave-KV achieves near-full-cache task quality with substantially lower memory use, improving average score and retrieval accuracy over prior compressed caches while reaching 8.3x KV memory reduction and 72.8 tokens per second at 64K context.

Recent years have seen multivariate time series foundation models (TSFMs) predominantly pretrained on synthetic data, which fails to capture real-world complexities. This paper introduces RMISC, a large-scale corpus with about 200 datasets and 142 billion time points across diverse domains. By pretraining four advanced TSFMs on univariate, synthetic multivariate, and real-world multivariate data, experiments show that real-world multivariate data significantly improves zero-shot generalization for both univariate and multivariate TSFMs, providing crucial resources and empirical evidence for stronger models.

FreqDepthKV is an inference-time KV cache compression method that factorizes adjacent-layer KV states into shared low-frequency depth components and sparse high-frequency residuals. A lightweight online probe assigns attention heads to different compression modes adaptively. On long-context QA, retrieval, summarization, and code generation benchmarks, it achieves 58.3 Exact Match, 63.0 F1, 32.5 ROUGE-L, and 48.1 pass@1 under a 32k-token prefill, with 70.4 tokens/s decoding throughput, 2.06s TTFT, and 6.2 GB peak KV memory—a 3.9× effective compression ratio, matching full KV performance and outperforming prior compressed-cache methods.

Multi-hop retrieval-augmented generation (RAG) requires sequential evidence acquisition, but existing methods typically rely on fixed pipelines or predefined control topologies, lacking learnable control over evidence operations. DynaKRAG formulates evidence acquisition as state-conditioned control, using a learned controller to select atomic operations (retrieval, query reformulation, evidence critique, sufficiency judging) and dynamically update the evidence state. Based on Qwen2.5-7B-Instruct, DynaKRAG achieves state-of-the-art F1 scores on three benchmarks including HotpotQA, demonstrating the benefit of coordinating retrieval, diagnosis, and gap-directed acquisition under evolving evidence state.

This paper presents Pitwall, a production system that generates natural-language Formula 1 strategy briefings in English, Spanish, and Portuguese, treating faithfulness as an architectural property rather than an aspiration. Every published sentence is decomposed into typed factual claims (positions, gaps, tyres, pace, overtakes, race control) and each claim is verified against the probabilistic race state that prompted it. The same verifier gates the fine-tuning data: of 3,045 model-written targets, only the 81.9% whose every claim is state-supported are retained, the rest falling back to a provably faithful template. The system is built on a vectorized Monte Carlo engine (N=2,000 per-lap race continuations) calibrated on 126 races (2018-2024) and validated on fully held-out 2025-2026 seasons (winner-in-top-3 90.3% over 155 backtests; held-out Brier 0.0745). End-to-end operation was confirmed at two consecutive live Grands Prix (Austria and Britain, 2026); at Silverstone a timestamped probability trace committed to disk before the outcome was known locked onto the eventual winner ten laps before the flag.

Simon Willison has released sqlite-migrate 0.2, a lightweight database migration system for SQLite built on top of his earlier sqlite-utils tool. This version simplifies the creation, application, and rollback of migration files, supporting incremental changes and version tracking. It is particularly suitable for small projects, prototyping, and embedded scenarios, offering a practical and easily integrated addition for developers managing SQLite schema evolution.