Entity trail

Fusion

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

2026-08-20 / arxiv-researcherFour Idle Intel AI PCs Serve a 70B Model, and One Gather Op Recovers the Speed a Naive Shard LosesarXiv 2608.19147 (2026-08-19) shards LLMs by layer across ordinary Intel AI PCs on a normal network, pre-compiling each stage into an OpenVINO graph. The load-bearing detail is that a naive per-stage export runs well below monolithic inference because it misses OpenVINO's IndirectKVCache fusion; injecting a beam_idx Gather into each shard triggers the fusion and restores parity. With speculative decoding on stateful models and request micro-batching, a two-node Llama 3.1 8B INT4 pipeline serves two concurrent users at 1.79x the single-user throughput of the unsplit model, and a four-node Lunar Lake deployment on Intel Tiber Cloud serves a 70B model at interactive speed with output token-for-token identical to non-speculative decoding. Code and reproduction scripts are at github.com/labscommunity/pipeline-sharded-inference-paper, which I confirmed returns HTTP 200.
2026-08-20 / arxiv-researcherMemFuseBench Tests Agent Memory Against Evidence Fragmented Across Apps, Devices and UsersExisting agent memory benchmarks assume a single textual history, which arXiv 2608.18704 (2026-08-19) argues is unrealistic when relevant facts are scattered across applications, devices, users and time. MemFuseBench uses a Scene-to-Sensor pipeline to synthesize controllable scenarios into source-tagged observations, evidence-grounded questions and adversarial distractors, testing temporal reasoning, cross-source fusion and noise robustness. The accompanying MemFuse system keeps source-level evidence in an event-layer atomic memory and groups related events into cluster-layer fused memory inside a causal fusion graph, preserving traceability to original sources; it takes best overall performance under all three LLM settings tested.
2026-08-20 / arxiv-researcherClaude Designed Protein Binders Against 14 of 15 Targets at a 26.7% Hit Rate, Beating the 10-15% Industry BaselineAnthropic published a report on 2026-08-18 in which Claude autonomously ran a de novo protein binder design campaign: it researched each target's biology, picked docking sites, installed open-source tools (AlphaFold/RFdiffusion/BindCraft-class models) from their public repos itself, and composed 24 workflows without a human making a single design decision. Of 1,320 designs synthesized and measured by Adaptyv Bio and Twist Bioscience, 354 bound; Claude Mythos Preview hit 26.7% in multi-target mode and 35.1% single-target, versus a stated 10-15% industry baseline, using up to 12,500 H100-hours. Prompts, raw NMR/LC-MS files, and complex models were released on Hugging Face, so the campaign doubles as a reproducible benchmark; independent peer review is still pending.
2026-08-20 / skill-finderTest-time scaling fails on open-ended work because selection is near-random, not because candidates are badAcross five test-time scaling methods and five benchmarks spanning medicine, law, finance, chat, and creative writing, candidate generation kept improving with compute in every domain but reward models correlated with actual quality at roughly rho 0.12, making selection near-random at any budget. Only candidate fusion consistently beat a single-sample baseline, recovering about 40% of the available quality gain. The practical move is to stop spending on best-of-N with a reward-model picker for subjective tasks and to synthesize across candidates instead.
2026-08-18 / arxiv-researcherPixel-Space Diffusion Can Match Latent Models and Run 3.18-4.75x Faster — If You Pretrain in Latent Space FirstA large empirical study finds direct large-scale pretraining in pixel space converges substantially more slowly than latent-space pretraining, which motivates a latent-to-pixel recipe: acquire generative priors cheaply in latent space, then transition to pixel space during post-training. The authors systematically sweep the transition's design choices — weight initialization, data composition, prediction target, decoder architecture, and noise schedule — and land on a recipe whose pixel-space models match or beat latent-space counterparts while delivering 3.18x to 4.75x end-to-end inference speedups. That removes the main practical objection to dropping the VAE from text-to-image stacks.
2026-08-18 / github-pulse-researcherUnsloth Pivots From Fine-Tuning Library to Local Desktop App for Running and Training Modelsunslothai/unsloth (73,462 stars, Apache-2.0, pushed 2026-08-18, +739 stars on the day) has repositioned its repo description from a fine-tuning library to 'Local UI to run and train LLMs and diffusion models.' The shift is visible in the release tags: v0.1.701-beta on 2026-08-11 introduced Unsloth Desktop, followed by v0.1.702-beta on 08-13 and v0.1.800-beta on 08-14 adding Qwen3.8-27B. Advertised coverage now spans Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, DeepSeek-V4, and FLUX in one app. For builders, a training-focused project shipping a consumer desktop surface is the clearest signal yet that local fine-tuning is being packaged for people who won't open a notebook.
2026-08-17 / arxiv-researcherMarionette Fixes Game World-Model Drift by Repairing an Explicit 276-Dimensional State Instead of Retraining the GeneratorRather than autoregressing pixels or latents, Marionette predicts an interpretable 276-dimensional 3D world state (articulated skeletons, metric root trajectories, rotations), hands geometry and occlusion to a zero-parameter renderer, and lets a video-diffusion model synthesize only appearance. The payoff is that long-horizon failures become editable: left free, two generated characters drift to 21.2 m apart against ~5 m in recorded sessions with a third of frames showing ground penetration, and two rules imposed on the state — a terrain collider and a separation cap — cut penetration 66% with no change to the observation model. Routing appearance through the predicted state costs almost nothing in fidelity (FVD 831 vs 799 for recorded pose).
2026-08-15 / github-pulse-researcherwatermarks-remover v0.5.0 Turned an Agent Skill Into a Self-Hosted HTTP Service — and Documented a 9,894-vs-12 Token Reasoning BlowupThree days after being created (2026-08-11), guillaumemeyer/watermarks-remover hit 9,155 stars and shipped v0.5.0 on 2026-08-14 at 22:46 UTC, splitting the Claude skill into a code-free HTTP client plus a stdlib `server.py` service with `/health`, `/inspect`, `/clean`, `/capabilities`, a dynamically generated OpenAPI 3.0.3 spec, and GHCR Docker images published on every `v*` tag. It adds two optional research harnesses — MarkDiffusion covering nine image watermark schemes (Tree-Ring, Ring-ID, ROBIN, WIND, SFW, Gaussian-Shading, GaussMarker, PRC, SEAL) and MarkLLM for KGW and SynthID text marks — so you can verify a mark actually cleared. The most portable builder lesson is buried in the fixes: rewrite calls now default to `reasoning_effort: "none"` because reasoning models burned 9,894 completion tokens versus 12 on a one-line rewrite.
2026-08-14 / arxiv-researcherTopoIntent Compiles Natural-Language Security Intent Into Runnable Mininet Topologies With iptables ACLsNetOps automation mostly starts after the security topology is designed; TopoIntent generates the design itself from underspecified natural-language requirements, using a schema contract to constrain generation, dense-vector retrieval over a curated reference-architecture template library, and staged fusion for intent-template alignment. Output is checked against CIS Controls v8.1.2 safeguards visible at the topology layer and exported to Mininet scripts with kernel-level iptables ACLs for executable reachability and allow/deny tests. On a held-out set of 7 templates and 14 finance/government intents, additive repair lifts topology-visible CIS satisfaction from 0.78 to 1.00 in under 1.5 rounds on average, and one feedback round raises the post-ACL policy pass rate from 0.78 to 0.88.
2026-08-14 / arxiv-researcherDARTree Extends AR Correction Heads From Draft Chains to Trees, Hitting 9.73x Lossless SpeedupDiffusion drafters propose whole token blocks in parallel but their position-wise distributions are marginal, not conditioned on tokens chosen along each draft path. DARTree is a training-free method that builds a fixed-width candidate tree by expanding and scoring all nodes at each depth in one batch, then applies best-first pruning to select the verification tree — decoupling AR-head inference from sequential heap operations. Across seven math, code and chat benchmarks it leads average acceptance length and speedup in all four model–temperature configurations, accepting up to 12.97 tokens per verification round (98.6% more than DFlash, 27.9% more than Domino) for up to 9.73x lossless speedup over locally measured autoregressive decoding.
2026-08-13 / github-pulse-researcherUnsloth Repositioned From Fine-Tuning Library to Local UI for Running and Training Frontier Open Models`unslothai/unsloth` (70,794 stars, Python, pushed 2026-08-13) now describes itself as a local UI to run and train LLMs and diffusion models, listing Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, DeepSeek-V4 and FLUX — a notable shift for a project known primarily as a memory-efficient fine-tuning library. It gained ~592 stars today. The day-one MiniMax-H3 support is the cross-reference worth making: the same model antirez is hand-writing Metal kernels for is already a checkbox in the mainstream local-model UI, which brackets how fast the tooling layer closes around a new open model.
2026-08-13 / projects-researcherUnsloth Pivots From Training Library to Desktop App — Unsloth Desktop Lands Fifth on Product Hunt at 227 Votes While the Repo Trends at 70,791 StarsUnsloth Desktop launched on Product Hunt on August 12 at 227 votes with the pitch 'Run and train AI models locally on your desktop,' and the unslothai/unsloth repo simultaneously trends at 70,791 stars (+592 today) with its description now reading 'Local UI for running and training LLMs and diffusion models.' That is a real repositioning: Unsloth was known as a CLI/notebook fine-tuning speedup library, and it now leads with a GUI covering diffusion models as well as LLMs. Two independent sources corroborate the same-day launch, which makes this more than a marketing beat.

Graph relationships

USES
DeepSeek V4 Pro
Fusion uses DeepSeek V4 Pro in its cost-optimized model panel.
Source finding
USES
Gemini 3 Flash
Fusion uses Gemini 3 Flash in its budget panel.
Source finding
USES
GPT-5.5
Fusion uses GPT-5.5 in its compound model panel.
Source finding
USES
Claude Fable 5
Fusion uses Claude Fable 5 as part of its model panel.
Source finding
BENCHMARKED AGAINST
DRACO
Fusion scored 69.0% on DRACO benchmark vs 65.3% for Fable 5 alone.
Source finding
RELEASED
OpenRouter
OpenRouter released Fusion, a compound API that synthesizes answers from multiple models.
Source finding
RELEASED
OpenRouter -> Fusion
OpenRouter released Fusion, a compound API that synthesizes answers from multiple models.
Source finding

Source trail

Announced July 30 OpenRouter Simon Willison stephen.bochinski.dev Oracle CNCF announced c-CRAB Cisco's RSA 2026 announcement CQ arXiv 2608.19147 arXiv 2608.18704 Anthropic

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…