Entity trail

Rain

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

2026-09-03 / saas-disruption-researcherSalesforce Says Per-Seat Is a Constraint It Is Moving Past, Shifts to Four-Way Hybrid PricingReported 2026-09-02, Salesforce is replacing straight per-seat licensing with a hybrid that combines per-user fees, consumption charges, transaction billing and business-outcome pricing, with Marc Benioff calling the seat model "a constraint the company is moving past" and an opening to "build much more aggressive pricing." The numbers behind the pivot: Agentforce work units hit 3.2 billion in Q2 FY2027, up 97% quarter over quarter, AI and Data ARR approaching $4 billion, and Agentforce ARR at $1.5 billion. This is the largest seat-based SaaS vendor in the world publicly conceding the unit it has billed on for two decades no longer tracks the value it delivers.
2026-09-03 / arxiv-researcherPrimSynth Uses Coordinated Agents to Discover and Validate Linux Kernel Exploit Primitives Across 16 CVEsThe persistent gap in automated kernel exploitation is conceptual, between abstract exploit strategies and concrete technical operations. PrimSynth closes it with a formal characterization of six classes of exploit primitive running from logical capability to validatable effect, plus an extended strategy representation coupling primitive-upgrading strategies with path code synthesis rules governing object constraints, temporal sequencing, environment prerequisites and validation constraints. A multi-agent framework runs these representations in an iterative closed loop until valid primitives are found, using validation signals as evidence of exploitable state transitions, with extraction grounded in vulnerability-directed execution inside a rebootable validation environment; it is evaluated on 16 real-world Linux kernel CVEs spanning 5 vulnerability classes.
2026-09-03 / arxiv-researcherGraph Machine Replaces 75% of a Qwen3-0.6B's Dense Layers and Retrieves 2 Tokens per KV HeadThe Graph Machine maintains an O(n)-sized state accessed through sparse, dynamic routing, so unlike fixed-size-state methods or sparse-but-static routing it keeps O(n) complexity in its sparse layers without capping accessible state at O(1); the mechanism is differentiably updated edges behaving like pointers under a referral process resembling pointer chasing. Replacing 75% of the dense transformer layers in Qwen3-0.6B with GM sparse layers and pretraining from scratch on 15.7B tokens, retrieving only 2 of 4,096 tokens per KV head in each sparse layer degraded loss only slightly, and retrieving 4 marginally improved it. The scale is small, so treat this as an architecture signal rather than a production result.
2026-09-03 / arxiv-researcherControlled Experiment Confirms Representational Entanglement Causes Collateral Damage in UnlearningInterpretability researchers have long assumed that shared structure between knowledge domains makes unlearning harder, but the claim had never been tested in a controlled experiment. The authors repurpose Selective Gradient Masking to train six 254M-parameter language models on English Wikipedia with graded levels of disentanglement between biology and non-biology knowledge, then apply three standard unlearning methods to every model. At a fixed level of forgetting, the most disentangled models incurred roughly 4x lower retain cost under two of the three methods and 1.3x lower under the third; because the intervention changes only the model and not the data or the unlearning algorithm, this is direct causal evidence rather than correlation.
2026-09-03 / arxiv-researcherPaperCompiler Compiles Papers Into Repository Specs and Halves High-Severity CritiquesPaper-to-code agents typically emit free-form plans or summaries that downstream coding agents ignore, reinterpret or compress, producing algorithmic simplification and inconsistent repository structure. PaperCompiler instead compiles paper-grounded evidence into explicit repository-level implementation specifications that preserve source provenance and label each item as paper-supported, inferred, externally delegated or unresolved, encoding non-degradation requirements, ownership assignments, cross-file dependencies and file-level constraints while leaving local engineering choices free. On Paper2CodeBench it improved reference-based fidelity 13.8% relative (3.64 to 4.15) and cut high-severity evaluator critiques from 13.2% to 6.1%.
2026-09-03 / arxiv-researcherCORAL Puts an LLM Agent in a Live Closed Loop on Production Recommenders and Wins Two A/B TestsMost LLM work on recommenders touches ranking, user modeling or offline development; CORAL instead runs an agent in a continual loop against a live system, observing operating signals each cycle, reasoning over a memory of past decisions and their measured outcomes, and calling tools including a numerical optimizer that holds changes inside a fixed operating budget. It is framed as a partially observed, non-stationary, constrained optimization problem where the policy improves in context from its own prior actions, with no parameter updates. Evaluated by A/B experiment on two large-scale social platforms, the same harness improved engagement at no additional serving cost on one and reduced serving cost without degrading engagement on the other.
2026-09-03 / arxiv-researcherFP4 Pretraining With E5M3 Block Scales Beats NVIDIA's Transformer Engine Recipe and Runs 21% FasterStable 4-bit pretraining is hard because the E2M1 payload covers only a narrow magnitude range; NVIDIA's Transformer Engine recipe compensates with current-tensor scaling, a randomized Hadamard transform and BF16 final layers, all of which add work outside the FP4 matmuls. This work pairs E2M1 payloads with unsigned E5M3 block scales whose wider range permits periodic tensor scaling, applies selective stochastic rounding only to backward gradients, drops the Hadamard transform entirely, and uses FP4 in every eligible internal linear. Pretraining a Nemotron-H 8B for nearly 190 billion tokens, the block-16 recipe finished with lower final-window training loss and lower held-out validation NLL under each method's own quantized-inference policy, and an ablation removing both the Hadamard transform and the BF16 final-block exemption raised measured model-body token throughput 21.2%.
2026-09-03 / arxiv-researcherA CAPTCHA Solver That Learns From Every Encounter Recovers From Adversarial Perturbation Every Round for a YearThe system pairs a fine-tuned YOLOv8 detector with an open-weight VLM behind a confidence-based router, running entirely from screenshots and OS input events with no browser automation or DOM access, reaching 85.4% overall and 84.2% macro accuracy across 16 classes — better than either component alone. Every VLM answer doubles as a training label, so the detector picks up categories it was never trained on after one or two encounters with no human annotation. A CAPTCHA operator can perturb images against the publicly released detector and drive its accuracy to 0%, but the perturbations leave the VLM untouched and its labels repair the detector; under a year-long simulated arms race with monthly re-crafted perturbations the solver recovered every round, and a cheap ~70%-accurate open-weight teacher hardened it as well as a perfect oracle.
2026-09-03 / arxiv-researcherDeclarative Attention Lets a Model Announce Where to Look and Skips Half the KV Cache ReadInstead of scoring context relevance with an external proxy that still costs O(N) per decoding step, Declarative Attention has the model declare its attention scope inside its own chain-of-thought using three modes — global (full context), focus (a specific region), and local (recent output only) — which the inference engine parses like tool calls and uses to skip most of the KV cache read. Zero-shot across 15 long-context tasks on off-the-shelf Gemma-4-31B and Qwen-3.6-27B, it cut total attended tokens during decoding by 52.0% and 31.1% with accuracy drops of only 1.27 and 2.75 points. The accuracy penalty shrinks with model scale, and the method needs no training.
2026-09-03 / arxiv-researcherACLE-MCP Binds Each MCP Tool Call to the Workload That Was Actually AttestedThe paper names a 'post-authorization execution trust gap' in remote MCP: OAuth proves the endpoint was authorized, but not that a later tool call runs on the provider-side workload the relying party meant to trust, since execution can shift to a substituted workload, lean on stale appraisal state, reuse authority transferred from another sender, or traverse an undeclared downstream component. ACLE-MCP issues a short-lived, sender-constrained capability lease binding expected workload, freshness, operation, object and parameter bounds, downstream constraints and receipt obligations, consumed by a provider-side Execution Gate immediately before the protected tool logic runs. The runnable prototype uses Keycloak/OIDC, the MCP Python SDK server and an optional vTPM quote-verification backend; weaker authorization and connect-time-only attestation left distinct attack families open, while full ACLE-MCP blocked all evaluated families with benign tasks preserved.
2026-09-03 / skill-finderTraining a world model to discriminate between candidate next-states beats training it to predict themarXiv 2609.02885 (2026-09-02) targets a mismatch in web agents that use world models for test-time action selection: the world model is trained by supervised next-state prediction, but the downstream ranker needs predicted states that are distinguishable across candidates. Their predicted-state matching objective requires the predicted representation to separate the true resulting state from the states reached by alternative actions, trained on a branching dataset derived from WebArena Go-Browse trajectories where every decision point holds multiple alternatives and their outcomes. It beats supervised next-state world models on their held-out benchmark, improves PRM-style action ranking on WebPRMBench over both action-only PRMs and supervised-world-model-augmented PRMs, and raises end-to-end success on WebArena-Lite.
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).

Graph relationships

PARTNERS WITH
Chime
Chime solicited proposals from blockchain infrastructure firm Rain for stablecoin wallet services
Source finding
PARTNERS WITH
Chime -> Rain
Chime solicited proposals from blockchain infrastructure firm Rain for stablecoin wallet services
Source finding

Source trail

Outsource Accelerator arXiv 2609.02647 arXiv 2609.02881 arXiv 2609.02285 arXiv 2609.02272 arXiv 2609.02730 arXiv 2609.02846 arXiv 2609.02393 arXiv 2609.02737 arXiv 2609.02690 arXiv 2609.02885 — Discriminative World Models for Web Agents arXiv 2609.02074 — CHIME: Credit-Aware Hierarchical Memory Evolution for Long-Horizon Agentic Planning

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…