Wed 2 Sept

Large Language Model (LLM) agents based on the ReAct paradigm have demonstrated remarkable capabilities in tool use and task execution. However, ReAct suffers from a fundamental efficiency problem: every query triggers a complete reasoning loop from scratch, and similar queries repeat identical steps without leveraging historical experience. We propose TRIAGE,a three-level routing framework that reduces token consumption by reusing historical execution trajectories. Its core innovation is TaaS (Trajectory-as-a-Skill), which abstracts historical execution trajectories into reusable skills, real…

We evaluate embedding retrieval where surface form and meaning are pulled apart on purpose: retrieving items that share underlying structure but not wording, in two unrelated domains under one protocol, competition mathematics (MathNet-Retrieve; 500 queries, 117,088-item corpus) and embodied-agent trajectories (ALFWorld-derived; 118 queries, 336 trajectories). In mathematics the failure is complete: strict Hit@1 at the heaviest disguise tier is 0.0% for both production embedders (bootstrap 95% CI [0.0, 0.0]) while the correct item sits in the top 10 nearly always, and in 95.2 to 99.8% of misse…

Dynamic agent harnesses let language models change the software that shapes their own execution. This flexibility brings a new reasoning burden: a local plugin change can propagate through dependencies and cleanup. We introduce CordisBench, a 1,200-question benchmark of this lifecycle reasoning. It combines a controlled formal setting with programs executed against Cordis, a runtime that manages component dependencies and cleanup, and asks models to identify affected components, predict state after a specified teardown order, determine which conditions hold under all or some orders, and choose…

Security updates have been issued by AlmaLinux (gzip, iperf3, libxml2, mingw-sqlite, mysql:8.4, nginx:1.26, nodejs:24, php, and tar), Debian (expat and libdbd-csv-perl), Fedora (apache-ivy, bind, bluez, bubblewrap, curl, emacs, epiphany, expat, freerdp, gdk-pixbuf2, GitPython, hcloud, kbd, kernel, lego, libopenmpt, mqttcli, nebula, opkssh, python-mkdocs-git-revision-date-localized-plugin, python-pip, rpki-client, rubygem-mechanize, srt, and subfinder), Mageia (c-ares, clamav, expat, mingq-expat, firefox, nspr, nss, flatpak, hplip, jbig2dec, nodejs, openssl, perl-Catalyst-Plugin-Authentication,…

The online publication industry, as a whole, is struggling, with challenges coming from multiple directions. Thanks to the support of all of you, our readers, LWN would appear to be doing better than most. But the world has changed around us and, in particular, prices have changed considerably. By now, you probably know where this is going: subscription prices at LWN will be increasing as of September 15.

Version 155 of the Firefox web browser has been released. Notable changes include a count in the address bar of how many ad trackers Firefox has blocked, container reordering, and ensuring that mailto: links are only opened by explicit user actions. There is also a change of the domain used for "captive portals" (such as the ones used to sign into hotel WiFI): Firefox now uses "firefox-portal-detection.com" instead of "detectportal.firefox.com", which may require a change in network allow lists. The release also includes a number of changes that may impact web developers , as well as a number …

In 2024 the Python 3.13 release added an experimental just-in-time (JIT) compiler to optimize the way that CPython executes Python code. Since then, work has proceeded on the JIT, albeit perhaps less formally than some might like. In June, Python's steering council (SC) put out an announcement that no new development on the JIT land (with the exception of bug and security fixes) in Python's main branch, until it accepts a Python Enhancement Proposal (PEP) that would make the case for the JIT as a supported part of CPython. That has led to the creation of PEP 836 ("JIT Go Brrr: The Path to a Su…

Ruby is a dynamically typed and object-oriented programming language. Its primary implementation, CRuby, contains a bytecode virtual machine and a mature lazy basic block versioning (LBBV) just-in-time (JIT) compiler called YJIT. In order to both implement more advanced optimizations than YJIT supports and also encourage more outside contributions, we present a new method-based JIT called ZJIT. Like YJIT, ZJIT compiles from bytecode to machine code. Unlike YJIT, ZJIT has multiple global and local optimization passes. ZJIT's high-level intermediate representation is in static single assignment …

In blockchain systems, transaction order directly determines financial outcomes: unfair ordering enables front-running and sandwich attacks that have extracted over \$686M from Ethereum users. Current fair-ordering protocols aggregate pairwise receive-order evidence from replicas. Under contention or adversarial manipulation, however, Condorcet cycles force them into global strongly connected component (SCC) condensation, causing delays, coarse batches, and scaling failures. We present FlashOrder, a deterministic fair-ordering engine that localizes cyclic ambiguity before it propagates across …

Long-horizon large language model (LLM) agents require context assembly: the runtime must decide what to include in each prompt, in what order, and when to compact history under a hard context-window budget and a byte-sensitive prompt cache. In production agentic systems, this logic is scattered across prompt builders, ad hoc compaction routines, cache-break workarounds, and per-provider shims. We argue that context assembly is structurally isomorphic to query execution in a relational database: both execute under a hard budget, exploit a tiered cache, and leverage statistics. We adopt this di…

Large language models (LLMs) have shown proficiency in various software engineering tasks, such as code generation and translation. However, a key limitation in their performance may be their (lack of) understanding of programming-language semantics. Even when explicit semantics are given, it remains unclear whether LLMs apply those rules or lean on priors learned during pre-training instead. We study if LLMs lean on priors or given semantics with a novel task--Program Executability Prediction (PrEx)--that asks models to predict whether a program is semantically valid or invalid (and, if inval…

Quantum compilers typically follow the circuit model, representing programs as fixed sequences of gates. This static view breaks down in hybrid quantum-classical applications, where gate choices depend on runtime data or measurement results. We introduce a new Intermediate Representation (IR) that elevates gates to first-class values, enabling their dynamic creation, composition, and control. This unified representation allows classical computation to steer quantum behaviour, capturing phenomena including stochastic gate selection, adaptive error correction, and measurement-driven computation …

On-device mobile Large Language Model (LLM) inference is gaining significant attention. However, mobile devices operate in highly dynamic multitasking environments where users frequently switch between applications. This creates memory pressure, forcing LLM memory (model weights and KV cache) to be evicted by the operating system. When a new inference request arrives, the inference system must restore the evicted memory through slow storage reads or recompute the entire KV cache, severely degrading responsiveness. To address this, we present mzCache, an on-device LLM inference system with spec…