Friday, October 16, 2026

Paper

Intern-Decision returns structured choices in one multimodal model pass

Intern-Decision evaluates all fields in a JSON decision skeleton in one causal pass; its 4B model reaches 90.02% average accuracy, with 44.16 ms mean latency for a 289-token request on an RTX 4090. Direct decision outputs could cut latency in agent control loops.

Original · huggingface.co by Tong Zhu et al.15 upvotes Code · 786 stars

Paper

JevAgentHarness delegates tool choices using fitted confidence thresholds

The benchmark finds that isolated decision accuracy did not predict agent success and that untuned delegation reduced success by 19 points; the harness fits thresholds on the agent’s own trajectories. This is a useful warning and design pattern for replacing expensive LLM calls.

Original · huggingface.co by Kunlun Zhu et al.14 upvotes Code · 3 stars

Release

Nemotron 3.5 Super VL combines multimodal reasoning and agent tool use

NVIDIA describes a 120B-parameter MoE model activating 12B parameters per token, with tool use and text, image, document, and video understanding; it reports up to 5× throughput versus open models in its size category. The architecture and claimed throughput are relevant to teams choosing models for specialized agents.

Original · developer.nvidia.com by Michelle Horton11 min read

Essay

Pre-commit hooks can steer coding agents toward approved diagnostics

A pre-commit hook scans diffs for direct calls that bypass a diagnostics library, then blocks the commit and prompts the agent to use it. Enforcing conventions in tooling can be more reliable than maintaining instructions that drift.

Original · simonwillison.net

Release

Google adds folder membership and scoped remediation to VPC Service Controls

VPC Service Controls now supports folder-based perimeter membership, granular exceptions and guided remediation for access denials. These controls address the challenge of protecting data while allowing autonomous agents to operate across enterprise projects.

Original · cloud.google.com by Pratik Bhangale5 min read

Deep dive

An AI agent uses Windows scans to investigate an unsupported scanner

The author describes setting up a Windows virtual machine so an AI agent could run scans and capture data for a scanner with an unsupported chipset. That workflow illustrates how agents can interact with real hardware during driver reverse engineering.

Original · blogs.gnome.org 7 min read HN 159 pts · 138 comments

Paper

SCOPE assigns text-to-image RL updates to semantic prompt spans

SCOPE pools policy-log-probability shifts within semantic spans and reports 9.48% average relative gain on GeneVal and 7.62% on T2I-CompBench over vanilla GRPO. It shows how structured credit assignment can improve prompt optimization without extra image renders.

Original · huggingface.co by Yufei Jiao et al.33 upvotes Code · 1 stars

AI models produce matching decompilation but still need human cleanup

The decomp.dev team says newer models can use its tools with little guidance and produce decent results in clean, documented codebases, but still need human fixes for errors and questionable shortcuts. It offers a grounded view of where model-driven coding work still needs review.

Original · decomp.dev 4 min read Lobsters 60 pts · 14 comments

Benchmark

Frontier models trail human references on a UI creativity benchmark

The benchmark excerpt gives a human reference score of 92.3 and a top listed model score of 65.5 for design creativity. Engineers evaluating generated interfaces can use the result as a reminder that functional output and creative diversity are different goals.

Original · arcfu.com HN 163 pts · 143 comments

Deep dive

Gemini 4 Argon scores 87% on a 3D embodied-reasoning benchmark

Argon scores 87% on ERQA, which tests reasoning about physical and spatial problems in 3D, and is presented for use across the robotics lifecycle. The benchmark offers a concrete signal of model capability, though the excerpt gives few implementation details.

Original · deepmind.google 16 min read

Deep dive

Rust for Linux proposes an Ownable adapter for C-allocated objects

The proposed Ownable trait lets Rust code manage the cleanup responsibility for objects allocated in C, without requiring Rust to deallocate them. It illustrates the ownership boundaries needed to integrate memory-safe code into an existing kernel.

Original · lwn.net by daroc

Release

Ubuntu 26.10 adopts Rust coreutils and post-quantum package signing

Ubuntu 26.10 adds Rust-based coreutils, post-quantum-ready package signing, TPM-backed disk encryption support and kernel 7.3. The changes matter to engineers operating Linux platforms and evaluating security and hardware support.

Original · lists.ubuntu.com HN 68 pts · 95 comments

Deep dive

A double-headed stack allocator addresses memory demands in performance-critical code

The piece introduces bump allocation in the context of memory-heavy animation blending in a game. Allocation strategies can matter in performance-critical software, though the excerpt does not explain the allocator’s implementation.

Original · blog.demofox.org 8 min read HN 93 pts · 35 comments

Deep dive

Android vendor hooks keep device-specific kernel changes out of mainline

Android devices still carry many out-of-tree kernel changes, with vendor hooks the preferred way to make them. The discussion is useful context for engineers managing platform-specific kernel behavior and upstream divergence.

Original · lwn.net by corbet

GCC developers discuss policy and practical uses for LLMs

GCC has adopted a restrictive policy for LLM use, while developers Jeff Law and Kyrill Tkachov discussed how the tools have helped them. The example offers a grounded view of AI-assisted work in a mature compiler project, though the excerpt gives few specifics.

Original · lwn.net by corbet

Benchmark

DuckDB benchmarks NVIDIA Vera with full TPC-H at scale factor 1,000

The benchmark compares DuckDB on NVIDIA’s 88-core Vera CPU and an Intel Xeon 6 using the full TPC-H workload at scale factor 1,000; Vera has up to 1.2 TB/s of memory bandwidth. The excerpt provides setup details but not performance results.

Original · duckdb.org by Gábor Szárnyas3 min read

Postgres Weekly covers agent-memory results and a queue slowed by savepoints

The roundup reports that raw conversation retrieval combined with extracted facts beat either alone in a LongMemEval rerun, while per-job savepoints slowed a Postgres queue as workers increased. Both findings offer practical leads for engineers building agent memory or database-backed workloads.

Original · postgresweekly.com 4 min read

Deep dive

Iceberg tables are trees of JSON, Avro and Parquet files

The guide walks through Iceberg metadata, manifest lists, manifests, data files and delete files, showing how DuckDB can inspect each layer with SQL. Understanding that structure helps diagnose slow or incorrect table queries without treating the table as a black box.

Original · duckdb.org by The DuckDB team13 min read

Release

Stripe Data Pipeline adds real-time sync to Postgres

Stripe Data Pipeline can now keep Postgres databases current with Stripe data, replacing custom sync integrations. The change may simplify the data plumbing behind applications that depend on current payment and account state.

Original · stripe.com

Release

Rust 1.99.1 restores wcslen and fixes x86_64 FFI bool returns

The point release fixes missing wcslen symbols on some targets and a miscompilation of bool return values across x86_64 FFI. These regressions matter to teams shipping Rust across targets or relying on C interoperability.

Original · blog.rust-lang.org Lobsters 12 pts · 5 comments

Essay

Deno is winding down its runtime as its team joins Cloudflare

Deno's runtime will wind down over the next year, while JSR continues on Cloudflare infrastructure and the team works to merge Deno's celld into workerd. The shift offers a useful look at the direction of the JavaScript runtime ecosystem.

Original · vlt.io 6 min read HN 50 pts · 4 comments

Guide

C23 conventions use [[nodiscard]] to catch ignored drawing results

One convention marks drawing functions [[nodiscard]] because callers need the returned rectangle for later logic, while allowing explicit handling when the value is unused. It’s a practical example of using newer C features to prevent bugs in a maintained project.

Original · abner.page 4 min read Lobsters 41 pts · 9 comments

Guide

C string views pair a pointer with an explicit length

The library separates borrowed string views from owned strings, representing each as a data pointer and length instead of relying only on NUL termination. This is a compact approach to making C string handling more explicit about ownership and bounds.

Original · eliasebner.com 17 min read Lobsters 22 pts · 21 comments

Release

Forgejo 17.0 changes how repository links resolve in rendered documents

Forgejo 17.0 resolves absolute links to repository resources from the repository root, changing their previous instance-root behavior; the release also warns users to review breaking changes before upgrading. The link change is intended to improve compatibility with other Git forges and ease migrations.

Original · forgejo.org 7 min read Lobsters 27 pts · 1 comment