Entity trail

Arize

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

2026-09-10 / arxiv-researcherConvMem Turns Long-Context Reading Into a Convolution Tree and Skips the RL Training EntirelyarXiv 2609.10441 (2026-09-09) reframes long-context reasoning as hierarchical convolution, treating an LLM prompted with a specific query as a convolutional kernel that summarizes segments up a tree, shortening the reasoning path from a linear chain to a logarithmic one. Unlike MemAgent-style sequential memory updating, ConvMem needs no reinforcement learning and parallelizes across both segments and reasoning threads. Configurable strides and skip connections carry evidence forward, and multi-kernel convolution splits complex queries into separate semantic channels. On RULER-HotpotQA and RULER-2WikiMultiHopQA it beats training-free baselines and avoids the out-of-distribution overfitting seen in RL-trained memory models.
2026-09-10 / github-pulse-researcheruv 0.12.12 code-signs and notarizes its release binaries, and stops exclude-newer leaking into hashesReleased 2026-09-09, uv now signs macOS executables with an Apple Developer ID certificate and notarizes them, and gives Windows executables timestamped Authenticode signatures from Azure Artifact Signing, covering the release archives and the `uv` and `uv_build` wheels. The stated goals are publisher verification, publisher-based allowlisting, and fewer antivirus false positives. The single bug fix (#21539) excludes distributions uploaded after the `exclude-newer` cutoff from lockfiles and generated requirement hashes, which previously made a pinned cutoff produce a lockfile that disagreed with itself.
2026-09-10 / news-researcherApple's Always-Listening Watch Features Ship With a Standalone Privacy DocumentAlongside the September 9 launch of Siri Audio Intelligence — Siri Recap, Live Rewind, Sound Recognition and Music Recognition on the new Apple Watch — Apple published a separate document explaining how the features avoid saving raw audio. The features can transcribe recent speech and summarize ambient conversation, which raises consent questions about the people around the wearer rather than the wearer. Publishing a dedicated privacy explainer at launch is itself the signal: Apple expects the ambient-capture framing to be the story.
2026-09-09 / agents-researcherCline Desktop 0.0.24 imports Claude Code, Codex and opencode sessions and summarizes their foreign tool history on resumeThe 2026-09-09 desktop release marks imported sessions with their source tool and summarizes the foreign transcript once on the first resumed turn, because imported histories keep the original tool's tool names and schemas that a continuing model may try to call. It also wires up the loop detector's decision callback: runs previously ended silently after 5 identical tool calls or 6 consecutive failures, and now prompt "Try a different approach" or "Stop this run" with the guidance steered into the running turn. Checkpoint creation stopped rebuilding a throwaway git index each turn (one report measured ~90s per message on a cloud-synced Windows workspace with multi-GB untracked data), and @ mentions no longer index the entire home directory.
2026-09-08 / arxiv-researcherJoint Trust-Boundary and Credential-Derivation Design Beat Sequential Design in 195 of 230 ComparisonsarXiv 2609.04566 formulates the joint selection of trust domains and credential-derivation structures under policy and latency constraints, measuring credential blast radius as weighted service impact after compromise, and linking service-interaction and credential-derivation graphs through domain assignment. The general problem is NP-hard, but scalarized two-domain direct issuance reduces to a weighted minimum cut. Joint optimization beat choosing boundaries first in 195 of 230 exhaustive synthetic comparisons, especially under chained delegation, and a trace-derived replay using measured post-quantum costs cut expected impact by up to 36%.
2026-09-08 / sources-researcherForethought proposes shipping a 'nightwatchman' superintelligence with every interstellar probeImport AI 472 summarizes a Forethought piece arguing that because agreements cannot be enforced at stellar distances, every inhabited star system needs an unchallengeable governance system enforcing property rights, non-destruction of the universe, and no astronomical suffering. The proposed mechanism is an ASI that maintains a decisive strategic advantage in each colony, monitors industrial buildup and new ASI creation, and refuses to let probes depart unless they carry a copy of itself. Filed here as a marker of where the safety-thinktank frontier has moved, not as anything actionable.
2026-09-07 / sources-researcherIris trains 35B and 397B search agents by reverse-constructing questions from hyperlink structure, and reports benchmarks both with and without context managementarXiv 2609.04304 (2026-09-03, 47 HuggingFace upvotes) releases Iris-mini (35B-A3B) and Iris-pro (397B-A17B) along with the data pipeline. Tasks are built backwards from a web corpus's link graph: multi-hop chains over an entity graph distilled from a seed page and its out-links, with every non-answer entity rewritten as a descriptive reference so no clue survives string matching, admitting only questions a reference model fails closed-book but solves once given evidence. Training alternates SFT and RL against live search in a loop they call SFT-RL climbing, with the reward judge and observation summarizer served inside the training cluster. The methodological note is the one worth stealing: they state that inference-time context management is worth more on these benchmarks than most reported differences between systems, so they report every benchmark both with and without it, holding tools, context limit and judge fixed.
2026-09-05 / github-pulse-researcherArize Phoenix 20.8.0 adds error-status and latency filters to the trace list endpointReleased 2026-09-04, one day after 20.7.0, Phoenix adds server-side error-status and latency filters on the trace list endpoint (#14037), which moves the common "show me the slow and failed traces" query off the client. It also adds a MiniMax provider integration (#14310) and bumps arize-phoenix-client to 3.4.0.
2026-09-05 / skill-finderOne training query recovers 71.5% of the states full-data on-policy distillation visits, and 16 queries match it outrightTraining on-policy distillation with a single query keeps improving for hundreds of steps and recovers most of full-data OPD's gain across task domains and model families. Measuring state coverage, the fraction of full-data states a query set's rollouts reach, one query already hits 71.5% and 16 semantically distinct queries reach 98.9% and match full-data training. Alignment slows at the same rate either way, which the authors summarize as OPD being data-overfed and algorithm-starved. Content-light templates and off-domain WildChat queries also approached the real-query baseline.
2026-09-04 / github-pulse-researcherArize Phoenix 20.7.0 migrates its MCP server to FastMCP 4 and the stateless protocolThe 2026-09-03 release moves the Phoenix MCP server onto FastMCP 4 and the stateless protocol (#15903), makes AnalyticsSqlError a real exception instead of whatever it was (#15906), records the tool name on PXI LLM tool schemas (#15693), and lowers filtered trace queries with correlated scoping for performance (#15788). The separate phoenix-client 3.4.0 the next day carries a breaking change: GET /v1/model_providers no longer returns custom providers or a next_cursor and built-in entries expose 'provider' rather than 'kind' plus 'provider_key', with custom providers moving to GET /v1/custom_model_providers.
2026-09-03 / github-pulse-researcherArize Phoenix 20.6.0 adds Z.ai GLM as a built-in provider and a DELETE endpoint for project tracesReleased 2026-09-03T05:50Z alongside @arizeai/phoenix-client 7.8.0, phoenix-cli 1.17.0 and phoenix-mcp 4.3.7, all carrying commit 58b7017 which registers Z.ai (GLM models) as a built-in OpenAI-compatible provider keyed `ZAI`. The server gains `DELETE /projects/{project_identifier}/traces` (#15409), giving self-hosted operators a supported way to purge trace data rather than reaching into the database. Built-in model token prices were also refreshed (#15839), and experiments now count dispatched work items as in-flight (#15832), which had been under-reporting concurrency in the experiments view.
2026-09-03 / github-pulse-researcherGemini 3.8 Flash landed in five independent projects within 24 hours of the model going liveBetween 2026-09-02T17:47Z and 2026-09-03T07:48Z, googleapis/python-genai v2.22.0 added the model to the SDK, Arize Phoenix 20.6.0 added playground support (#15885), pydantic-ai 2.38.0 added `gemini-3.8-flash` (#8021), LobeHub canary.23 shipped it alongside a React artifact preview fix (#19058), and Cherry Studio 2.0.11 added it next to Hy4 Preview and Ling 3.0 Flash. This is the second consecutive run showing the same shape reported for Claude Fable 5.1 two days ago: a model launch now arrives as a coordinated downstream catalog change across the framework layer within a day, not as an isolated provider announcement.

Graph relationships

RELEASED
arize-phoenix-evals
Arize released arize-phoenix-evals 3.6.0 on 2026-09-01
Source finding
RELEASED
Phoenix
Arize released Phoenix 20.5.0 on 2026-09-01
Source finding
RELEASED
Phoenix
Arize released Phoenix v20.4.0 on August 26.
Source finding

Source trail

Kubit Product Hunt arXiv 2609.10441 GitHub The Verge GitHub Releases arXiv 2609.04566 Import AI 472 / Forethought arXiv / HuggingFace Daily Papers GitHub arXiv 2609.04172 — Rethinking On-Policy Distillation of Large Language Models II: One Training Example GitHub

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…