Nexus Speeds Up Agentic LLMs with Depth-Adaptive KV-Cache and Tool Routing
Nexus Speeds Up Agentic LLMs with Depth-Adaptive KV-Cache and Tool Routing
Nexus is a novel method for optimizing agentic LLMs in MCP environments, addressing the quadratic prefill bottleneck that prolongs time-to-first-token as tool registries grow. It employs depth-adaptive KV-cache splicing and decouples retrieval from tool routing on unified memory. This reduces overhead from re-encoding verbose tool schemas, enabling faster response times for agents that frequently call external APIs. Nexus is particularly effective for large-scale tool ecosystems, promising significant latency reductions in complex agent workflows.
Self-speculation speeds up reasoning models, reduces latency
An arXiv paper proposes self-speculation, where LLMs predict their own reasoning traces to speed up inference without external draft models, ideal for latency-sensitive applications.
Inhibitory Attention Reduces 'Lost-in-the-Middle' Errors in Long EHR Analysis
This paper addresses the ‘lost-in-the-middle’ effect in LLMs when processing electronic health records (EHRs), which often exceed 100,000 tokens. The proposed inhibitory attention mechanism characterizes and mitigates this issue by recalibrating attention weights. In clinical reasoning tasks, it significantly improves the retrieval of critical information located in the middle of long contexts. This work tackles a key barrier to deploying AI for medical diagnosis on complex, real-world patient data.
When Clean Data Hurts: New Work on Learning Under Monotone Corruptions
This paper investigates a critical flaw in the PAC learning model: the impact of monotone corruptions where correctly labeled samples from an unrelated or adversarial source are added to an i.i.d. training set. It demonstrates that such ‘clean’ yet non-i.i.d. data harms the performance of optimal learners that rely on the classic assumption. The study formalizes learning under monotone corruptions and analyzes its impact on generalization, warning that simply ensuring correct labels in datasets is insufficient.
New 'Wrong-Physics' Backdoor Attack Targets Neural PDE Operators in Solver Archives
Researchers introduced ‘cross-parameter relinking’, a new data-poisoning backdoor attack targeting neural PDE operators trained on reusable solver archives. By injecting carefully crafted data points that adhere to ‘wrong physics’, the attack compromises the model’s integrity. Standard validation methods, including prediction error and parameter-agnostic plausibility checks, often fail to detect these backdoors. As a result, models can produce catastrophically incorrect predictions for specific problems, posing a serious risk for scientific machine learning applications.
Multilingual Verifier Bias Found in RLVR, Hindering Non-English Math Reasoning
A new study challenges the assumption that answer verifiers in Reinforcement Learning with Verifiable Rewards (RLVR) are language-neutral. It reveals a significant multilingual verifier bias, identifying a ‘cross-lingual selection bottleneck’. Through benchmark tests and rollout diagnosis, the authors show that models struggle to select correct answers in different languages, leading to suboptimal training outcomes for non-English mathematical reasoning. This highlights a critical flaw in the standard recipe for training LLMs on multilingual reasoning tasks.
Vercel uses Turborepo cache hits to prevent build machine downgrades
Elastic build machines now consider Turborepo cache hits when deciding whether to use smaller machines, preventing unnecessary downgrades for warm-cache builds.
GitHub Accessibility Scanner plugin checks if alt text actually works
The GitHub Blog announces a plugin for the GitHub Accessibility Scanner to ensure alt text is actually accessible, helping developers catch issues automated checks might miss.
Linux trick: make a SQLite database directly executable
Farid Zakaria describes a Linux pattern for setting a SQLite file’s 4-byte application ID to SELF, making the database file directly executable.
llm-anthropic 0.27 adds compatibility with Anthropic Python library v1.0.0
llm-anthropic 0.27 is released, primarily providing compatibility with Anthropic v1.0.0 Python library, which switches from httpx to httpx2. OpenAI made the same change in v3.0.0 two weeks ago.
GPT-5.6 lands in Kiro with better price-performance for developers
OpenAI announces GPT-5.6 is now available in Kiro, helping developers plan, build, review, and test software with improved price-performance.
Claude Code v2.1.243 adds loop usage breakdown and model picker settings
Claude Code v2.1.243 adds a Loops breakdown to /usage (per-loop run count, total tokens, tokens per run, last run) for spotting runaway tasks, plus a modelPicker setting to curate the /model picker with ordered lists.
OpenClaw beta adds GPT-5.6 Sol, Terra, Luna, Ultra reasoning support
OpenClaw 2026.8.1-beta.3 adds GPT-5.6 Sol, Terra, Luna, and Ultra reasoning support across OpenClaw and Codex runtime, Control UI first-run setup improvements, Puppeteer-compatible CDP relay for Chrome sessions, and explicit external gateway support.
Vercel Sandbox goes global with four initial regions
Vercel Sandbox now runs globally with four initial regions: iad1 (Washington D.C.), sfo1 (San Francisco), cle1 (Cleveland), and cdg1 (Paris). iad1 remains default, with all Vercel regions coming soon.
OpenAI Codex releases 0.150.0-alpha.8
OpenAI Codex released version 0.150.0-alpha.8, a pre-release with multiple updates.