Search “Reasoning”

Microsoft and City University of Hong Kong propose Taste-Bench to evaluate agent decision-making

Researchers from Microsoft and City University of Hong Kong have built Taste-Bench to evaluate the long-term decision-making capabilities of AI agents on long tasks. The benchmark contains 502 binary-choice questions. According to The Decoder, the best-performing model, GPT-5.6 Sol, achieved an average accuracy of only 59.7%, while GPT-5.5 reached 59.5%. When key clues only become apparent after a large amount of subsequent work, the average accuracy of all models drops to 21%, close to random guessing, and increasing the reasoning budget brings almost no improvement. The researchers also distilled the teacher model's judgments using Qwen3.6-27B as the base model; the student's accuracy on unseen tasks rose from 30% to 47.9%, and its success rate on 41 SWE-bench Pro tasks increased from 14.6% to 33.7%.

Perplexity Decider launches on Vercel AI Gateway

On October 17, 2026, Perplexity's Decider V1.2 decision model launched on Vercel AI Gateway. Vercel developers can now use perplexity/pplx-decider-v1.2 for complex agentic workflows and perplexity/pplx-decider-flash-v1.2 for ultra-fast decisions. Aravind Srinivas reposted the news. Earlier the same day, reports said Perplexity had open-sourced two decision models, pplx-decider-flash-v1.2 and pplx-decider-v1.2: the former has 4B parameters and costs 1 cent per million tokens, and is called the strongest 4B model on the HF Decision Index; the latter has 27B parameters and costs 2 cents per million tokens, and ranks 2nd on the overall HF Decision Index leaderboard. Both are now available on the Perplexity Decisions API and offer open weights on Hugging Face. Perplexity Developers said this generation of open-weight multimodal decision models improves both text and vision scores, delivers inference up to 5 times faster, and completes all questions in a single request in one forward pass.

Hao AI Lab evaluates detective agent teams with 200 cases

Hao AI Lab evaluated three detective agent configurations on 200 Victorian-era London cases: the Jev cluster acting alone, GPT6-Astra acting alone, and GPT6-Astra leading the Jev cluster. The institution's published conclusion is that the cluster can cover the clues, but a reasoning leader is needed to spot contradictions and pursue them in depth. In 100 harder cases, with more paid witnesses than honest witnesses, GPT6-Astra plus the Jev cluster solved 93, GPT6-Astra alone solved 59, the Jev cluster alone solved 42, and a rule-based leader plus the Jev cluster solved only 0–1. There are currently no follow-up developments.

Hao AI Lab announced the results of its multi-agent crime-solving benchmark

On October 17, 2026, Hao AI Lab announced the results of its multi-agent crime-solving benchmark, which used 200 Victorian London detective cases. Among 100 harder cases in which paid witnesses outnumbered honest ones, GPT6-Astra leading the Jev cluster solved 93 cases, GPT6-Astra alone solved 59 cases, the Jev cluster alone solved 42 cases, and a rule-based leader plus the Jev cluster solved only 0–1 cases. Hao AI Lab also noted that when working alone, each of the two approaches has its own failure mode: the leaderless Jev swarm is like a rumor mill—one child misreads a clue, its companions copy the vote, and the whole street accepts the wrong answer; while the lone genius is limited by time and cannot deliver the decisive clue to the door. The authors conclude that, in the face of organized lies, a reasoning leader is needed who will ask, "Who is speaking?"

Nadella says new research shows machine learning can improve 2- to 6-week weather forecasts

Satya Nadella introduced new research showing that machine learning can improve weather forecasts two to six weeks ahead, a time scale that has historically been difficult to predict. The research has been published in Nature. Nadella said the work could help respond to extreme weather, manage natural resources, and support agricultural decision-making. He also offered a broader idea: the future lies not only in more capable models, but in building intelligent verification systems around models—learning where models systematically err, reasoning about causal relationships, and continuously calibrating predictions against reality. At present, this news has only this one source, and there are no subsequent developments.

OpenRouter launches RespanAI decision model Span-02

OpenRouter announced the launch of RespanAI's open-weight decision model Span-02. The model accepts text and image inputs and returns options, yes/no/unknown, or scores with probabilities in response to custom questions; pricing is $0.01 per million input tokens with free output, and the context length is 131K. OpenRouter later clarified that Span-02 runs on the OpenRouter Decisions API, uses the same request format as other decision models, and switching requires only changing one line with the model name; images and text are placed in the same state, with a maximum of 4 images per request. Respan previously introduced Span-02 as supporting anytime inference, returning the current best decision once a deadline is given, allowing configurable reasoning duration, and supporting multimodal viewing of images and screens. The model is now live on OpenRouter.

Databricks has been collecting npm malware samples since November 2025

Databricks has been collecting npm malware samples since November 2025. Its blog discloses that, among the 11,413 package-versions scanned and verified, 96.2% hid their payloads in source code or on external servers, while only 3.8% (432) bundled binaries; however, the proportion of bundled-binary packages from which no clear malicious conclusion could be drawn was more than twice that of pure source-code packages (14.4% versus 6.2%). The team evaluated 9 models, including Claude Haiku 4.5, Gemini 3.5 Flash, GPT-5.4-nano, Qwen3-Next 80B, and others, and injected precomputed static evidence as capability profiles during the inference stage to improve detection quality for both budget-tier and frontier models. Currently, public information is still based on what this blog discloses, and there have been no subsequent developments.

Aleph Alpha releases the Kolibri-1 open-weight model

On October 16, 2026, Aleph Alpha released the open-weight model Kolibri-1. According to heise online, the model has 78B total parameters and uses a MoE architecture with 384 experts and 6 routed experts, activating only 3.46B parameters during inference, with a context of 256K tokens that can be extrapolated to about 1 million tokens. The model focuses on German and English, and its tokenizer segments German words more token-efficiently; for example, the German Bundestag glossary requires only 1,137 tokens, whereas the compared DeepSeek-V4.1-Flash requires 1,910 tokens. Pretraining used 20 trillion tokens, with data up to June 18, 2026. The report's author tested it and said its answers were correct and its German reasoning process evident, and said the Merlin-Arthur protocol makes the model hallucinate less in benchmarks, but its actual performance still needs to be tested in real-world use. There are currently no subsequent developments.

Reflection announces a $10,000 Hub credit for the Beam project

Reflection announces a $10,000 Hub credit for the Beam project. The news was released on October 16, 2026. At the same time, Hugging Face released the Beam model: a 501B-parameter sparse MoE open-source model aimed at long-context reasoning and agent tasks with a 1M context, activating only about 23B parameters per token. The model uses 52 layers of hybrid attention, of which 43 layers use a 512-token sliding window and 9 layers use global attention; the KV cache grows by 36 KiB per token, about one-sixth that of a full-attention scheme. Official checkpoints are provided in BF16, FP8, and NVFP4; NVFP4 can compress the weights from about 1 TB to about 280 GB, and day-0 support is already available in transformers, vLLM, and SGLang. There are currently no subsequent developments.

English translation is being prepared.

English translation is being prepared.

Artificial Analysis publishes evaluation results for Reflection's first model, Beam

On October 16, 2026, Artificial Analysis published the evaluation results for Beam, the first model from Reflection: Beam scored 30 on the Artificial Analysis Intelligence Index, ranking second among US open-weight models, behind only Meta's Muse Spark Open (39) and ahead of NVIDIA's Nemotron 3.5 Super VL (26). The model is an open-weight model with 501B total parameters and 23B activated parameters, consuming about 26k output tokens per intelligence index task, of which about 19k are used for reasoning and 7k for answering; it scored 48.5% on AutomationBench-AA and only 3.0% on Terminal-Bench 4.0; AA-Omniscience accuracy was 33% with a hallucination rate of 47%. On October 17, Arena.ai announced that Beam-1.0-501B-A23B entered Agent Arena, ranking 12th among open-weight models with a Net Improvement of -3.67% and a margin of error of ±3.84%; on the Code Arena: WebDev open-source leaderboard it ranked 22nd with a score of 1438.

English translation is being prepared.

English translation is being prepared.

Chollet says Argon performs similarly to Astra on ARC-AGI-3

François Chollet commented that Gemini 4 Argon performs similarly to Astra on ARC-AGI-3, with Google returning to the frontier. On the same day, ARC Prize published the verified results for Gemini 4 Argon on the ARC-AGI series: ARC-AGI-3 scored 97.5% under the provider adapter harness at a cost of $4.8K; under the standard harness it scored 49.8% at a cost of $5.8K. ARC-AGI-2 scored 91.3% at $0.32 per task; ARC-AGI-1 scored 98.5% at $0.16 per task. ARC Prize said Google DeepMind is the second frontier AI lab to exceed 95% on ARC-AGI-3, and Argon's 97.5% is second only to GPT-6 Astra's 99.9%. The standard harness test could not complete all 55 games due to API issues; the high-reasoning setting completed only 19 games, so no score was published.

English translation is being prepared.

Elorian releases its first AI model, Glimpse

On October 17, 2026, Elorian released its first AI model, Glimpse. According to Testing Catalog, Glimpse can, during its thinking process, return to a specific point in a video, zoom in on details, or replay half a second of video at a higher frame rate before answering. Elorian's own testing shows that Glimpse outperforms GPT-6 Astra in reading scientific charts and describing short videos; in a simulated task of planning for a robotic arm to place cream cheese into a bowl, Glimpse cost $0.035, compared with $0.249 for GPT-6 Astra; in a demonstration, Glimpse solved 5 mazes of increasing difficulty, while Gemini 3.8 Flash solved 1/5. Another report the same day said that Elorian positions Glimpse as the first visual reasoning model, used to compare scene changes at different points in a video, and calls this frame-based reasoning visual thinking; during its reasoning process it can return to the source video, crop regions, extract specified frames, or run code to measure and compare what it sees, a mechanism called visual scratch space; co-founder Andrew M. Dai said it has frontier performance competitive with Astra on key public visual reasoning benchmarks. At present, there have been no further evidence-supported developments since the model's release.

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

Meta launches Muse Wonder and Muse Wonder Full models

On October 13, 2026, Meta released the Muse Wonder series, comprising two models, Muse Wonder and Muse Wonder Full, which it called its most capable models to date, achieving SOTA on personal and professional agent tasks, covering professional domains such as law, finance, and health as well as frontier coding capabilities. Both models feature Active Memory and Active Focus, and were trained, evaluated, and released according to the Meta Superintelligence Scaling Framework. According to Artificial Analysis data, the two models scored 52 and 56 respectively on Intelligence Index v4.3 at maximum reasoning effort, higher than the 48 of Muse Spark 1.3 (max); Muse Wonder is already available on Meta's first-party API and Muse Code, while Muse Wonder Full is available only to Muse Code users. On the same day, Arena added the two models to Agent Arena testing, and Simon Willison released the llm-meta-ai 0.2 plugin supporting the new models. On October 14, Arena announced Muse W