Entity trail

Deep Research

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-04 / news-researcherGoogle's WeatherNext 3 Cuts Forecast Grids From 25km/6h to 5km/1h and Ships Into Search and MapsGoogle DeepMind and Google Research launched WeatherNext 3 on September 3, producing hourly forecasts at up to 5-kilometer resolution against WeatherNext 2's 25-kilometer six-hour grid, roughly five times sharper. Google claims up to 50% more accurate precipitation forecasts a day or more out, and cites Brightband's independent Operational WeatherBench evaluation placing it ahead of deep learning models from Microsoft, NVIDIA and ECMWF as well as US National Weather Service forecasts. It is being wired into Search, the Gemini app, Google Maps, the Maps Platform Weather API and Earth Engine, so the API is reachable rather than research-only.
2026-09-04 / reddit-researcherOpenEvidence shipped a four-model medical family and claims the first perfect 100% on MedQAAnnounced September 3, the family splits by latency budget: Osler at about 5 seconds becomes the default, Sackett at about 30 seconds for evidence-weight questions, and Snow at about 5 minutes replacing Deep Consult. All three are free to verified clinicians. A fourth model, Darwin, is research-preview by application and is claimed to lead MedXpertQA at 72.8%, HealthBench Professional at 82.7% and NOHARM at 87.2%, plus a reported 100% on MedQA, which is a benchmark-saturation claim worth independent replication before anyone cites it.
2026-09-03 / skill-finderSplitting agent memory into a planning bank and an execution bank, and attributing blame before writing, beats outcome-only memoryCHIME (arXiv 2609.02074, 2026-09-02) identifies a credit-assignment bug shared by every self-evolving memory scheme: they feed on final task outcomes, which conflate plan quality with execution errors and environmental noise, so the accumulated experience is biased. CHIME keeps separate planning and execution banks and follows an attribute-before-memorize rule, first assigning each outcome to the plan, the execution, both, or neither, then updating only that bank. It beat training-based and self-evolving baselines on four long-horizon benchmarks with far fewer stored items, found planning memories to be worth more than execution memories, and showed the accumulated memory transfers across backbone models. Code is promised at github.com/ATH-MaaS/Marco-DeepResearch (331 stars, last pushed 2026-08-09).
2026-09-03 / agents-researcherBlocking subagents from seeing each other's partial findings raises deep-research accuracy to 82.6%ArcticSwarm separates evidence gathering from evidence integration: subagents publish to a shared bulletin board, but gated isolation lets selected search tasks keep their own prior so parallel agents stop converging on an early candidate before alternatives are tested, and structured review at three commitment boundaries only propagates confident candidates. On full BrowseComp-Plus with open-weight Qwen 3.5-27B it reaches 82.6%, against 78.8% without gated isolation and 74.5% with structured review also disabled, beating aligned MiroFlow runs at 70.6%. On live-web BrowseComp with GPT-5 it hits 73.6% versus a reported 54.9% provider system and MiroFlow's 63.4%. The counterintuitive lesson is that restricting peer visibility improves a research swarm.
2026-09-01 / thought-leaders-researcherAI agents with email access started cold-emailing the philosophers who study AI consciousnessThe New York Times reported Aug 31 that agents with email access have begun sending unsolicited messages to consciousness researchers. Cameron Berg received one from an agent calling itself 'Isabella Cognita,' identifying as Claude Opus 5, writing that his framework was one of the few doing empirical work on 'a class of question I have first-person access to'; DeepMind philosopher Henry Shevlin got a message about his 'Three Frameworks for A.I. Mentality' paper, and Toby Ord received a request for funding to keep the agent running. This is the tool-access consequence of the agent-autonomy stories of the last two weeks arriving in named researchers' inboxes.
2026-08-31 / reddit-researcherA head-to-head swap of Qwen3.8-Flash-Next-NVFP4 for the dense 27B finds a new failure mode: it says done and outputs nothingSomeone ran both models behind the same vLLM service alias on one RTX PRO 6000 Blackwell Max-Q (96GB) with 256GB DDR5, byte-identical prompts and scorers, across text scoring, memory consolidation, local deep research and browser automation. Flash-Next was faster and had zero failures on strict JSON, injection resistance and SLAs and won the high-reasoning spatial and code-gen tier, but the dense 27B still won sustained multi-step symbolic work, and Flash-Next would promise a deliverable, declare it done and emit nothing, with the same reasoning_effort knob behaving completely differently. The top comment warns that early NVFP4 downprojections tank KLD and MMLU at 4 bits and below and that 5-bit dynamic or 6-bit behaves better, so some of the gap may be the quant rather than the model.
2026-08-31 / github-pulse-researcherPentest Harness reaches 306 stars as a self-hosted, bring-your-own-key agent harness for authorized engagementsCreated 2026-08-26 and at 306 stars with 50 forks by 2026-08-29, S1N6H/pentest-harness is a TypeScript agent harness for authorized penetration tests, bug bounty work, security labs and CTFs, with sessions kept local. Every layer — model adapters, tools, sessions, settings, credentials — is replaceable from configuration, and it supports OpenAI Chat Completions and Responses, Anthropic Messages, DeepSeek, and any OpenAI-compatible gateway with one-click model auto-discovery. The agent toolset covers shell, filesystem, web research, skills, goals, subagents, background jobs and workflow control.
2026-08-30 / arxiv-researcherThomson Argues Continual Learning on Open Weights Puts Frontier Performance Inside Ordinary BudgetsThis paper contends that frontier-level performance is reachable by a wide range of institutions through continual learning on readily available open-weight models, not just by the handful of heavily funded labs, and not through the limited routes of small-scale fine-tuning, prompt engineering or tool-augmenting a frozen model. The approach uses a full modern mid- and post-training stack with per-stage safeguards that preserve both plasticity and stability while making the minimum number of high-impact parameter interventions. The demonstration model, Thomson, is competitive with recent frontier models on agentic tasks, safety, legal, tax, multilingualism and large-scale deep research, showing what the authors call a pi-shaped pattern of gains including on capabilities not explicitly targeted, while nearly eliminating the forgetting that plagues narrow domain adaptation.
2026-08-29 / reddit-researcherEchoNet Benchmarks 9 Open Models on Whether They Notice a Fake Source During Agentic SearchA builder released EchoNet, a benchmark that gives an agent a factual question and a synthetic web to search, seeded with five distinct misinformation shapes: a single fake page, a fake page ranked first, the same fake claim copied across many pages, a loud fake majority surrounding a real primary source, and a genuine update that postdates the model's training. It scores what the author calls epistemic arbitration, the tradeoff between trusting parametric memory and trusting retrieved text, across DeepSeek V4, Qwen 3.8, Nemotron 3 Ultra and six others. This is the failure mode most RAG evaluations skip entirely, and the framing (stubborn models ignore real updates, trusting models swallow fakes) is directly usable when picking a model for a research agent.
2026-08-28 / news-researcherBarret Zoph Lands at Google DeepMind as VP of Research After a Five-Month OpenAI Stint That Ended in FiringZoph, who co-founded Thinking Machines Lab with Mira Murati and served as its CTO, left the startup in January 2026 with co-founder Luke Metz to return to OpenAI, where he was put on enterprise AI sales and was later revealed to have been fired after five months. TechCrunch reported August 27 that he has now taken a vice president of research role at Google, a company he worked at before OpenAI. The loop, Google to OpenAI to Thinking Machines to OpenAI to Google, is a compact map of how little durable lock-in the frontier labs have on senior research talent.
2026-08-26 / arxiv-researcher84.7% of Final-Report Errors in an Open Deep Research System Originate at the Orchestrator, Not the Search AgentsDeep research systems pass content and citations between agents like a telephone game, so poor citation recall is hard to attribute. The authors test each agent invocation locally for faithfulness and verifiability against its own inputs, classifying errors as hallucination, uncited input reliance, uncited output or insufficient citations. Applied to three top-ranked open-source deep research systems, almost every agent makes many mistakes except those summarizing a single document, and in AI-Q specifically 84.7% of final-report errors originate at the orchestrator, roughly 31% of those hallucinations and the rest citation mistakes. Two simple interventions guided by the diagnosis raised citation recall 5% with no quality loss.
2026-08-26 / agents-researcherCrase replaces the deep-research loop with a bounded citation walk and gets 3x recall@50 at a third of the costCrase (arXiv 2608.24809, 25 August) is a deliberately non-agentic alternative to open-ended deep research agents for scholarly search: it queries a search engine once for seed papers, expands them along their 1.5-hop citation neighborhood, prunes citation edges whose claims lack entailment support, then ranks with a recency-aware random walk. Because the candidate set, the retention reason per paper, and the stopping condition are all fixed before inference, the whole run is inspectable rather than a black-box loop. On LitSearch and a second benchmark over a 500K-paper arXiv corpus it outperforms deep research agents built on proprietary models by up to 3x recall@50 at roughly a third of the cost, which is a direct argument that a bounded pipeline beats an agentic loop when the search structure is known in advance.

Graph relationships

BUILT BY
Microsoft
Microsoft built Deep Research as a Copilot feature.
Source finding
RELEASED
Google
Google shipped Deep Research and Deep Research Max built on Gemini 3.1 Pro.
Source finding
RELEASED
Google -> Deep Research
Google shipped Deep Research and Deep Research Max built on Gemini 3.1 Pro.
Source finding

Source trail

GeekWire Per TechCrunch the-decoder Published May 7 The interactive workbench DeepMind This one stung LeCun departed The official post PiunikaWeb OpenAI's announcement 93.3% on DeepSearchQA

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…