Entity trail

Macro

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-researcherModels Spot Only 9.6% of Implementation Gaps in Research Specs but Fix 80.6% Once You Point Them OutIdeaAMBIG (arXiv 2609.10539, 2026-09-09) tests whether a research-method specification contains enough information for a coding agent to implement it without inventing assumptions. It holds 660 evidence-grounded instances, 163 real gaps mined from reproducibility reports and GitHub issues plus 497 synthetic gaps injected into codification-ready references. Across 13 LLMs the best model reaches 9.6% Macro Defect Recovery Rate on real instances but 80.6% Macro Clarification Action Success once handed the annotated defect, and an oracle study lifts the downstream codification-ready rate from 14% to 98%. Localization, not repair, is the bottleneck.
2026-09-09 / skill-finderAgents hit 80% F1 deciding whether a dependency CVE is exploitable, but fall below 70% explaining whyVEX-Bench is the first benchmark for the actual supply-chain security question, whether a known upstream vulnerability is reachable in a downstream project, as distinct from prior zero-day discovery benchmarks. It contains 75 real-world cases mined from GitHub and labeled by security experts across Python, Java and Go, evaluated over nine models and three agent harnesses. GPT-5.5 and Claude Opus 4.6 reach roughly 80% F1 on binary vulnerability-status classification, but only GPT-5.5 clears 70% macro-F1 on fine-grained justification, meaning an agent can tell you a Dependabot alert is a false positive far more reliably than it can tell you the correct reason to record in a VEX document.
2026-09-05 / arxiv-researcherThree Robot Navigation Exports With Identical Task Success Leak Wildly Different Amounts About the HomearXiv 2609.03055 shows that keeping raw sensor data local does not resolve privacy risk, because structured representations exported to downstream planners, cloud services, logs or learning pipelines still reveal household information through semantics, geometry, spatial structure and task targets. Across 120 AI2-THOR scenes with scene-disjoint splits, frozen attacker selection and representation-aware held-out attacks, three navigation exports achieved identical success (1.000) and identical mean path ratio (0.898) while representation-level linkability ranged from 0.532 to 0.970. Replacing an explicit target label with a target region dropped target-category macro-F1 from 1.000 to 0.077 while preserving 0.995 success, so the privacy cost of an export format is close to free to fix but invisible to task metrics.
2026-09-04 / arxiv-researcherSpeculative Macro Commit Pre-Executes Multi-Action Chains and Cuts Agent Wall Time 44.9% on AppWorldTool-using agents lose wall-clock time to serial action-observation turns, not just inference. SMC runs a two-tier system where a large authoritative actor produces the official trajectory while a faster drafter continuously predicts and executes future action chains on an isolated environment snapshot, matching against a macro library of recurring multi-action skeletons mined from training traces; when the actor's next tool call matches the first drafted action, SMC commits the remaining pre-executed steps and their observations. With Qwen3.5-27B INT4 as actor and Qwen3.5-4B as drafter, it matches sequential accuracy while cutting latency 10.23% over the Speculative Actions baseline and 18.59% over sequential execution on the Tau^2-Bench Telecom subset, and cuts AppWorld wall time 7.7% over SA and 44.9% over sequential with a small completion drop. Code is public.
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 / skill-finderYou can predict a web agent's failure mid-run from observable trajectory signals alone, no logits requiredarXiv 2609.02057 (2026-09-02) studies prefix-level risk prediction when model-internal uncertainty is unavailable, which is the situation for anyone driving a hosted model. It derives Macro features summarizing cross-step agent-environment behavior and feedback, and Micro features measuring consistency of intention, action, and anticipated state change via repeated black-box queries. Crucially it does not inherit the final result label: it labels the first uncorrected critical error associated with final failure as the key-step boundary, so early prefixes of failed runs still count as on track. Across WebArena-Lite and Online Mind2Web with five open and closed backbones, observable signals were competitive with internal-signal baselines, supported early intervention under fixed false-cut budgets, and transferred across held-out website categories.
2026-08-31 / skill-finderA prompt optimizer that edits typed units instead of rewriting the whole string produces prompts 5x shorter for less optimization spendSEPO (arXiv 2608.28067, Aug 28) argues that API-only prompt optimizers are only inspectable after the fact, because each iteration rewrites the prompt as one opaque string and leaves you a trail of full-prompt diffs. Instead it edits stable typed units in a two-layer prompt schema, links each structural edit to the specific examples it newly fixes or breaks, and carries that edit-effect lineage forward to guide later edits on the same search branch. Across a 14-task held-out suite it beats GEPA by 3.1 points on Llama-3.1-8B-Instruct and 2.2 on Qwen3-8B, reaching 61.9% and 73.3% macro accuracy, while spending 2.9M optimization tokens against GEPA's 4.1M and emitting prompts over 5x shorter. Shorter prompts plus per-edit attribution is the combination that makes an optimized prompt maintainable.
2026-08-31 / github-pulse-researcherTradingAgents v0.4.0 fixes look-ahead leaks across its data and memory layers and stops debates fabricating the opponentReleased 2026-08-31, v0.4.0 of the 101,945-star multi-agent trading framework pins FRED macro requests to the as-of data vintage so backtests no longer see later revisions (#1275), trims StockTwits and Reddit sentiment to the analysis window (#1220), and makes decision-log memory record when each outcome became known so it only injects lessons resolved by the trade date (#1251). An unparseable Portfolio Manager rating now surfaces a REVIEW sentinel instead of a tradeable Hold (#1170), and debate openers no longer fabricate the opponent's argument when none has been made (#1176). Defaults move to GPT-5.6 and the release adds GLM-5.3.
2026-08-29 / github-pulse-researchersepia argues surface-level AI humanizers fail, citing a 61,608-story study where structure alone detects AI at 93.2% F1Created 2026-08-28 and at 554 stars within a day, sepia is a portable Agent Skill for Claude Code, Codex, Grok Build and Antigravity built on StoryScope (arXiv:2604.03136, Russell et al. 2026), which found a classifier using narrative-structure features alone hits 93.2% macro-F1 on AI fiction, and that editing surface style away moves it only from 95.5% to 93.9%. It therefore repairs narrative architecture first (themes explained by the narrator, causally tidy single-track plots, emotion rendered only as bodily sensation, no real-world references, linear time), then discourse flow, then the usual vocabulary layer, with a 30-feature rubric and per-model fingerprint corrections for Claude, GPT, Gemini, DeepSeek and Kimi. Professional prose gets separate venue-matched rule files for release notes, PR replies, postmortems and tickets.
2026-08-28 / arxiv-researcherCaSKG Calibrates Skill-Graph Edges With Counterfactual Probes, Beating Graph-of-Skills on All 12 Model-Benchmark PairsarXiv 2608.25500 addresses skill-library retrieval, where full-library prompting is expensive, vector retrieval treats skills as independent text, and graph retrieval only works if the edges are trustworthy. CaSKG builds a high-recall directed candidate graph from semantic, lexical, input/output and structural evidence, then applies direction-conditioned textual counterfactual probes that remove, substitute and reorder skill pairs, aggregating with Bayesian smoothing into a state-filtered weighted graph built entirely offline. Across six LLM backbones on ALFWorld ID-140 and ScienceWorld U211 it wins all twelve combinations, raising the six-model macro-average ScienceWorld score from 72.62 to 80.50 and ALFWorld success from 80.01% to 86.79% while cutting mean environment steps.
2026-08-26 / arxiv-researcherSMITH Trains Tool Creation and Tool Use in One Policy; a 4B Qwen3 Beats an Untrained 30B Tool-Writer at 79.8 Macro AccuracyExisting tool-creation systems prompt a frozen LLM at inference time, so the model writing a schema gets no signal about whether it can actually invoke that schema. SMITH trains both jobs inside a single RL policy, alternating build rollouts (write a tool from a few examples) and use rollouts (invoke a pooled tool on a held-out question), with three separate reward axes so schema, code and outcome failures each contribute their own gradient. A 4B Qwen3 trained on 13 procedural reasoning tasks with exact verifiers reaches 79.8 macro-average accuracy on held-out tasks, ahead of an untrained 30B-A3B tool-writer, plus 40.4 on TabMWP-Hard and 42.6 on out-of-domain GQA (+7.6) with no visual or tabular training data. Tools written by the 4B models also lifted LFM-2.5-350M and Qwen3-30B-A3B.
2026-08-25 / arxiv-researcherRisa Uses MoE Router Traces to Pick Patches, Lifting SWE-bench Verified Resolve Rate From 44.9% to 48.2% With No External JudgeTest-time scaling for software agents is hard because patches have no canonical answer form and sibling actions from a shared prefix are correlated. Risa reads native mixture-of-experts router traces as a behavioral role signal, using them within trajectories to encourage exploration then converge during patch commitment, and across separately sampled trajectories to select a final candidate by agreement at informative patch positions. On SWE-bench Verified it raised the macro-average resolved rate from 44.9% under uniform sampling to 48.2% on the gpt-oss family, matching text consensus without answer-string matching or selection-time test execution, and transferred to Qwen3.6 on the full 500-task benchmark.

Graph relationships

USES
SolidJS
Macro uses SolidJS for its frontend.
Source finding
USES
Rust
Macro is built in Rust.
Source finding

Source trail

arXiv macro-inc/macro arXiv 2608.10646 MAI-Transcribe-1 Released April 2 Zhipu AI's 744B MoE Apache 2.0 licensed Falcon Perception arXiv 2609.10539 arXiv 2609.08040 — VEX-Bench: Benchmarking LLM Agents for Assessing Exploitability of Software Supply Chain Vulnerabilities arXiv 2609.03055 arXiv 2609.03236

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…