All Articles · SURL AI News

AgenticGenPlan Generates Robot Planning Programs Without LLM Calls During Evaluation

The study tests reusable control programs in 28 simulated environment types. In the subset with classical planner comparisons, the best method reaches a 95% average success rate. Results are limited to fully observable states; reproducing them requires checking black-box access controls, evaluation seeds, and upfront synthesis costs.

2026-09-26

AI 研究

RECLAIM tests research agents on 100 papers; full-materials group reaches a 41% reproduction rate

The benchmark asks agents to install dependencies, debug issues, and run experiments themselves, then has an independent model audit the execution evidence. Public code and traces let developers trace points of failure, but the results remain limited by model coverage and single-run testing.

2026-09-25

GitHub Repo

langchain-anthropic 1.7.4 Supports Adding and Removing Tools Mid-Conversation

The new version uses system messages to convey tool changes, allowing agents to adjust their capabilities in stages and potentially preserve existing prompt caches. The feature is still subject to model, message-position, and cache-setting limitations, and no performance test data is available.

2026-09-25

GitHub Repo

langchain-fireworks 1.6.3 Wraps Invalid Historical Arguments to Fix Agent Conversation Resumption

The new release wraps malformed or non-object JSON from historical tool calls in diagnostic data, preventing a subsequent request from being rejected because of stale arguments. The change preserves call-and-result pairing; actual task recovery rates remain unverified.

2026-09-25

GitHub Repo

TensorFlow 2.22 Release Candidate Adds Tiered Batch Waiting and a Separate Warmup Thread Pool

BatchFunction adds configurable wait times by request criticality and a control for handling warmup work in a separate thread pool. No performance benchmarks have been published; deployments still need to validate scheduler settings and tail latency.

2026-09-25

GitHub Repo

Hermes Agent 0.21.5 Adds Per-Profile Start and Stop Controls for Partial Shared-Gateway Maintenance

The release lets admins independently stop and start a selected agent on a shared gateway, with a persistent marker to retain its disabled state. Already-dispatched scheduled jobs continue running, and compatibility should be checked for older multi-process deployments.

2026-09-25

GitHub Repo

Ultralytics 8.4.162 adds image batch prefetching to overlap loading with CUDA inference

The new release uses a background thread to preload the next batch of images, reducing the time the GPU spends waiting for data. Published benchmarks cover specific YOLO26n configurations, while the release summary and merge history still differ on the feature’s scope.

2026-09-25

GitHub Repo

AutoGPT 0.8.1 Centralizes Sandbox Connections, Adds Optional TCP Egress Proxy

The release centralizes E2B sandbox creation, reconnection, and proxy credential rotation. The feature is disabled by default and requires an external proxy; DNS and QUIC traffic fall outside this TCP channel.

2026-09-25

AI 研究

CliffCompaction releases agent compaction package that trims context using verbatim excerpts

The package deletes or truncates older content through an API proxy, bringing rule-based compaction to existing coding agents. The paper’s cost comparison uses an idealized caching model, and the published defaults differ from the experimental settings.

2026-09-25

AI 研究

WISE reuses attention regions in recurrent models, with a measured 1.36× speedup in the attention pipeline

The method retains all recurrent steps, using early exploration to narrow the scope of subsequent attention computation. Performance figures cover only attention workloads; the quality drop at 4K context length and the public reproduction materials still require verification.

2026-09-25

GitHub Repo

langchain-openai 1.6.6 fixes silent streaming errors, raises exceptions for failure events

The release propagates explicit failure events in Responses API streams to the application layer, with tests added for both synchronous and asynchronous interfaces. Truncated streams without terminal events still require separate verification, and existing exception handling and retry strategies need review.

2026-09-25

推論系統

Inferact open-sources TPU inference megakernel, with Kimi K3 exceeding 700 tokens per second in low-concurrency decoding

The implementation combines 92 layers in a single Pallas call, improving memory bandwidth utilization through cross-layer weight prefetching. The official comparison is limited to a specific topology and small batches, and speculative decoding performance also depends on the draft acceptance rate.

2026-09-24