All Articles · SURL AI News

RRSI adds search constraints to agent self-modification, with tests showing gains across benchmarks

The study limits the scope of changes and uses evaluation variability and inference costs to screen modifications to agent frameworks. The public implementation enables replication, but the scale of token savings depends on the baseline.

2026-09-23

代理工具

Strands Introduces a General-Purpose Agent Harness with Context Offloading and Prompt Caching

Strands harness combines tools, memory, and context management into ready-to-deploy agents, with Python and TypeScript interfaces. The team has published cost comparisons, but actual benefits still depend on the task, model, and caching conditions.

2026-09-23

NVIDIA unveils eight-speaker diarization model with minimum recommended buffering latency of 320 milliseconds

Nemotron 3 Diarization uses 100 million parameters to process streaming and recorded audio, retaining a speaker cache to keep labels consistent across segments. The latency figures cover only audio buffering; actual deployments still need to measure speech recognition, transmission, and post-processing costs.

2026-09-23

GPT-6 Sol and Luna Launch, Cutting Token Prices and Expanding Agent Cache Controls

Sol’s standard input and output prices are half its predecessor’s promotional rates, while Luna also gets a price cut. Cache updates let agents retain prefixes while adjusting reasoning and tool use, but write rates and task retries still affect total costs.

2026-09-23

推論系統

ComputeArena publishes local inference benchmarking workflow with signed reports documenting models and environments

The tool compares BaseRT and llama.cpp using fixed workloads, retaining file hashes and per-run timings. Differences in warm-up behavior across runtimes and the reliability of self-reported data still limit how the leaderboard can be interpreted.

2026-09-23

AI 研究

Extending Recursive Model Inference to 512 Steps Raises Cumulative Sudoku Solve Rates

By extending computation with the same weights, researchers observed that some initially unsuccessful puzzles were solved later. The findings offer clues for studying stopping criteria, but do not yet establish cost-effectiveness in deployed services.

2026-09-23

AI 研究

GPU Topology Study Adjusts Memory Allocation, Measures 14.3% Gain in H200 Decode Throughput

The study incorporates compute unit placement and memory partitions into inference configurations, improving decode throughput under a specific Qwen3-8B workload. The prototype requires driver modifications, and the results await code release and reproduction across environments.

2026-09-23

AI 研究

Agent Self-Correction Study Adds External Validation, Revealing Regression Risks from Rule Updates

The study found that some corrective rules improved earlier failures but broke previously successful cases. The public implementation enables re-evaluation, but its division of roles differs from the paper, so effectiveness and costs still need to be checked against specific versions.

2026-09-23

AI 研究

AISI to add inference-compute evaluations to Evaluation Cards, publishing settings for comparisons across reports

AISI and EvalEval use a shared format to organize results and settings from existing research, covering five main experimental benchmarks and related cybersecurity evaluations. The data helps trace budgets and feedback conditions, but format validation does not amount to independent replication.

2026-09-23

GitHub Repo

vLLM Reportedly Terminates Engine After External KV Cache Read Failure; Scope Remains Unclear

A report involving a development build and a hybrid-attention model shows that a single request failure was followed by a fatal scheduler assertion. The documented failure policy should contain failures at the request level, but no public fix or cross-environment validation is available yet.

2026-09-22

Claude Opus 5.5 lowers token prices and introduces multiple breaking changes to the agent API

The new version reduces input, output, and cache read prices and makes adaptive thinking always active. Forced tool calls, edits to conversation history, and streamed progress displays all require renewed validation.

2026-09-22

GitHub Repo

vLLM publishes Qwen3.8-2.4T disaggregated inference benchmarks, reaching about 5,000 total tokens per second per GPU

The team benchmarked separate prefill and decode configurations on GB300 NVL72, providing version details and a configuration repository. Peak throughput and roughly 180 generated tokens per second per user come from different configurations and still require validation under real workloads.

2026-09-22