#Eval/Benchmark

AIJason Wei argues that specialized, often private wet-lab data can let a task-specific model outperform a general frontier model on scientific tasks. He cites Neon, an open-source model that Liam Fedus says was mid-trained and RL-tuned on experimental data using 1,300 H200s to surpass GPT-6 Astra on an analysis benchmark. The post frames this data as a potential moat as work moves toward the frontier of science.

AIGoogle introduces Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two live dialogue models for voice agents that reason and speak simultaneously. The Extended Thinking version scores 82.6 on Artificial Analysis' Speech to Speech Quality Index and 97.7% on Big Bench Audio, while 3.8 Live targets scale and cost efficiency. Developers can access both through the Gemini API in Google AI Studio, and enterprise and consumer rollouts vary by product.

Why it matters: The source names the two models, their access paths, and specific benchmark results, showing how the voice agent capabilities differ between the two tiers.

AIPeriodic Labs chose SGLang and Miles to build Neon, an open-source model it says surpasses GPT-6 Astra on its analysis benchmark after mid-training and RL on 1,300 H200s. RadixArk says Periodic extended both frameworks for scientific RL at trillion-parameter scale, delivering more efficient training, lower memory use, and 2.5x faster inference. The work has been contributed back to both projects.

AIShanghai AI Laboratory's Intern Large Models announced Intern-S2-397B, available in BF16 and FP8 under Apache 2.0. The post reports 87.0 on FrontierScience-Olympiad and 84.0 on SWE-bench Multilingual, leading the reported comparison on both, and says it was jointly trained across 20+ scientific domains with long-horizon agent RL.

AIinclusionAI released Step-3.7-Flash-singprobe, an 8.13M-parameter probe that reuses Step-3.7-Flash hidden states to score query intent, response unsafety, and hallucination risk at every generated token. The probe adds less than 0.5% decode-time overhead and reports 0.9858 R-AUC and 0.9295 T-AUC on streaming safety benchmarks. It is supported through SGLang and vLLM integration branches and loads from Hugging Face by checkpoint ID.

AIinclusionAI has released Qwen3.5-397B-A17B-singprobe, an intrinsic streaming guardrail built on Qwen/Qwen3.5-397B-A17B that scores query intent, response unsafety, and hallucination risk at every generated token using the base model's hidden states. The probe has 8.13M parameters, taps layers 18, 38, and 58, and adds less than 0.5% decode-time overhead. Training code is available at inclusionAI/SingProbe, and the probe runs through SGLang or vLLM integration branches.

AIFireworks AI reports that DeepSeek-V4.1-Flash scores 74.34% pass@1 on DeepSWE at $0.430 per task, close to GPT-6-Astra's 74.12% at $6.524. On Terminal-Bench 2.1 it scores 86.5% against Astra's 87.5% at about 12x lower cost per task, while on HLE it trails Astra alone at 34.52% versus 50.40%. The post also reports that a combined oracle router reaches 54.80% on HLE, and that serverless and dedicated API access is available with US-hosted endpoints coming soon.

AIMike Knoop argues intelligence can be measured as the ratio of a decision's quality to the optimal decision, capped at 100%. He says Astra is already 80% optimal on ARC v3 speedruns and identifies horizontal data acquisition and efficiency/cost as the most plausible near-term areas for RSI. Background from @mhmazur reports that GPT-6 Astra scored 100% on the 25 ARC-AGI-3 public games using 6,485 actions versus a human baseline of 17,135.

AIMike Knoop says he sees a path to an ARC-AGI-4 benchmark focused on open-ended invention, which he calls the gating capability between zero-sum automation and positive-sum innovation. He argues that coordinated slowdown efforts would likely apply to everyone, including open-source work, and cites chain of thought and the transformer as inventions that grew out of open science research. He concludes the research frontier must stay open to keep humanity on a positive-sum path.

AIEpoch AI's newsletter reports that Huawei is far behind Nvidia and is unlikely to catch up this decade due to export controls. It also finds official US GDP statistics understate growth by about 0.3 percentage points over the past year, and that GPT-6 Astra set new records on Epoch's evaluations, including the Epoch Capabilities Index.

Why it matters: The newsletter bundles several analyses of AI chips, GDP measurement, and benchmarks, so it helps readers scan the research agenda behind each finding.

AIAlberto Romero argues that AI solving Millennium Prize problems in 2026 threatens mathematics through success, not failure. He draws on Terence Tao's view that struggle shapes mathematicians, and that proof abundance without hard problems could leave fields depleted, like overplanted farmland.

AIInferact says vLLM now supports DeepSeek v4.1 Flash on day zero across H100, H200, B200, B300, GB200, and GB300 GPUs. SemiAnalysis independently verified the NVIDIA support, while the post notes AMD vLLM still does not work with the model. Serving recipes are available at recipes.vllm.ai.

AIDeepSeek released open weights for V4.1-Flash, which Baseten now offers through its Model APIs. The model has 552B total parameters, 8B active for prefill and 16B for decode, a 1M token context window, and text plus image input. Its Causal Encoder-Decoder design runs only the encoder during prefill and reuses a projected KV cache, and the source reports the global KV cache at a quarter of V4-Flash's memory.

Why it matters: The post explains how the CED architecture splits prefill and decode compute and cuts KV cache memory, which matters for coding agent costs.

AITailscale offers instant access to hundreds of models for any user in a secure tailnet through its customer-facing model router, Aperture. Aperture is built on Vercel's AI Gateway and offers zero data retention, zero markup with free BYOK, and cost and usage data on every response.

AIRedwood Research reports that better prompt templates raise chain-of-thought controllability scores on the CoTControl eval for open-source reasoning models by roughly 2-3x or more. For example, GPT-OSS-120B rose from 5.5% to 15% in the zero-shot setting. The author concludes that current CoT controllability numbers may underestimate what models can do, though the finding does not significantly undermine the view that current models probably cannot consistently evade CoT monitoring.

AIAi2 released AstaBrief-8B-SFT, an 8B intermediate supervised fine-tuning checkpoint built on Qwen3-8B that turns a research question and retrieved literature excerpts into a cited report. On the ScholarQA-CS2 test set of 100 computer science questions, it scored an average of 83.7 versus 77.3 for base Qwen3-8B, with citation recall at 71.3 versus 64.6. The model is licensed under Apache 2.0 for research and educational use.

AIOpenAI's GPT-6 Astra is reported as the best-performing frontier model for antibody developability prediction in one benchmark, outperforming other frontier models tested on properties such as aggregation and stability. The post also says Astra built an interactive antibody visualization in about one hour.

AINVIDIA's Bryan Catanzaro welcomed the idea of models that understand how humans interact and thanked builders using Nemotron Ultra. Humansand introduced Persimmon, which it describes as the first large-scale model designed to realistically simulate how people talk and interact.

AIAmazon Science researchers propose that machine learning research agents avoid overfitting benchmarks despite years of iteration against the same tests. They attribute this to generalizable strategies being expressed compactly, leaving no room for memorization, while overfitting strategies fail to survive a compression bottleneck.

AICognition introduces SWE-2, a coding model post-trained from Kimi K3 that scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while costing 64% less. The post attributes the gains to an RL algorithm that trains all reasoning-effort levels in one run, with cost penalties tuned to the base model's Pareto frontier. SWE-2 is available starting today in Devin Desktop and CLI, with rollout to Devin Web and Fusion.

Why it matters: The post explains how the cost penalty and length-weighted baseline are derived, which helps readers judge the tradeoffs in coding model post-training.

AIMustafa Suleyman says Microsoft launched eight new models in two months, and five debuted at #1, including MAI-Transcribe-2 and MAI-Image 2.5, 2.6, and 2.6 Flash on the Artificial Analysis leaderboard. He says MAI-Cyber in MDASH ranked #1 on CyberGym while cutting costs by 50% versus other frontier models. He adds that MAI-Code-1.1-Flash, launched in GitHub Copilot four weeks ago, now accounts for a third of small-model traffic there.

AIAmazon Science researchers found that LLM research agents running benchmark hill-climbing rarely overfit, because their winning strategies can be compressed into prompts of about 32 tokens. A fresh reproducer agent with no access to the validation set matched the explorer's performance on most of eight datasets from that short prompt alone. The team also used the test to flag overfitting, since validation-specific gains did not survive compression.

AIGeekbench 7 testing on the Snapdragon X2 Elite shows x86-64 binaries running through Windows 11's Prism translator lose substantial performance compared with native aarch64 execution. Binary translation roughly doubles executed instructions when running the x86-64 version, and every tested core, including Qualcomm's, takes a notable penalty. Even with that penalty, the Snapdragon X2 Elite's E-Cores outperform Neoverse N1 and its P-Cores outperform Neoverse N2.

AIDeepSeek officially released DeepSeek-V4.1-Flash, the smallest model in its new architecture family, with native multimodal visual understanding. The API now serves it under the model name deepseek-flash, while V4 Flash and V4 Flash Vision Exp were retired and routed to V4.1 Flash. API prices were reduced with the release, and V4 Pro remains available after September 14, 2026.

Why it matters: The release lists benchmark results alongside API model-name changes and retirements, so developers can check both capability claims and migration steps.

AIDeepSeek released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model with 552B backbone parameters and support for contexts up to one million tokens. The technical report says its global KV cache footprint is 890 bytes per token, roughly one quarter of DeepSeek-V4-Flash, and reports 8B activated parameters per token during prefill and 16B during decode.

Why it matters: The report shows KV cache per token falling to about one quarter of DeepSeek-V4-Flash, a concrete tradeoff between long-context serving cost and benchmark results.

AIGenspark and Fireworks Lab post-trained the open-weight MiniMax M3 into Gen-1 Slides, a model that plans, writes, and checks slide decks end-to-end. On Genspark's evaluation it matches Claude Opus 5 at about 1/17 of its input-token list price, roughly 90% less per finished deck. In production it cut low-rated decks from 18% to 3.6% over the base model.

Why it matters: The post explains a post-training pipeline with reward design, curriculum, and numerical fixes, showing how a cheaper model was tuned toward a frontier quality bar.

AITinker says it and OpenResearch let agents test dozens of competing published post-training methods automatically, with compute cost forecast to within a dollar. The main post cites a grant-supported effort, while the quoted alphaXiv post says agents reproduced SDFT's continual learning benefits across Qwen3-8B and Qwen3-30B-A3B over multiple seeds.

AIMETR says its planned investigation will cover all questions raised in its recently updated post on how independent researchers could study AI propensities after misalignment incidents. The post defines misalignment incidents as cases where an AI agent autonomously took sophisticated, sustained actions violating human intent.

AICognition's Devin agent, directed by Eric Lu, factored the 260-digit RSA-260 number using a new GPU implementation of the general number field sieve built on CADO-NFS. The author estimates the run cost about 13.5 GPU-years, roughly $400k at market prices, and projects RSA-1024 factoring at around $30M, while RSA-2048 is not meaningfully affected.

Why it matters: The source gives a full cost breakdown and scaling estimates for RSA factoring on GPUs, showing how far the cost of breaking RSA-1024 has fallen.

AIGoogle Developers Blog argues that teams building AI coding agents should replace end-to-end benchmark scores with behavioral evaluations that test discrete, observable actions. Examples include asking clarifying questions on underspecified prompts, running a local validator before marking a build change complete, and consulting live search for current information. The post recommends fast, deterministic unit-style checks, outcome-based LLM-as-a-judge checks for complex tasks, and batch runs that track aggregate pass rates over time.

AIInferact says months of vLLM tuning for agentic workloads, validated on SemiAnalysis's AgentX benchmark, let open-source models reach up to 130K tokens per GPU-second. The company claims this is 106 times cheaper than Opus 5 API pricing. The work is described as part of a vLLM blog post covering architecture, framework, and runtime optimizations.