Entity trail

Arena

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

2026-08-21 / reddit-researcherQwen3.8-27B at FP8 With xhigh Reasoning Scored 29/30 on AIME 2026, Matching BF16 and Beating It at Medium EffortA practitioner benchmarked Qwen3.8-27B on the MathArena/aime_2026 dataset across BF16 and FP8 weights at medium and xhigh reasoning effort. FP8 at xhigh beat BF16 at medium and equalled BF16 at xhigh while running faster, which is the practically useful result: the quantization costs nothing on this task, the reasoning budget is what matters. On problem 7 both BF16 and FP8 at xhigh exhausted the generation budget without emitting an answer, so that miss is a context-limit failure rather than a wrong answer.
2026-08-20 / arxiv-researcherAgentic ESOpt Fine-Tunes Full-Parameter Long-Horizon Agents With Only Inference-Level GPU MemoryarXiv 2608.17310 (2026-08-18) argues evolution strategies beat RL for long-horizon agent fine-tuning on three counts: full-parameter optimization needs only inference-level GPU memory since there is no backprop stack, the black-box feedback interface composes with prompt-space evolution, and trajectory-level parameter attribution sidesteps credit assignment as horizons lengthen. Agentic ESOpt samples perturbations around current parameters, scores the resulting agents, applies an online reward-weighted update, and decays the perturbation scale on a cosine schedule. Full-parameter optimization of Qwen-3.5-27B improves the no-skill baseline by 6.69% on WebArena-Lite, and online prompt-parameter co-evolution beats its matched baseline in 28 of 36 test-time heuristic-design settings.
2026-08-18 / arxiv-researcherPrior Labs Open-Sources RelArena-α, TabPFN-Rel and RPI — and Flattening Still Beats Specialized Relational ArchitecturesPrior Labs released three Apache-2.0 pieces: RelArena-α, a unified framework standardizing data loading, evaluation, and tuning for RelBench v1's 21 entity-level forecasting tasks (inspired by TabArena); TabPFN-Rel, a relational harness for TabPFN-3 that improves on RDBLearn via deep feature synthesis with per-task join depth d ∈ {2,3,4}; and RPI, a model-agnostic YAML-driven interface that turns CSV or Parquet files into a task in two lines of code, shipped with an agent skill and example specs for all 21 tasks. TabPFN-Rel currently ranks first on RelArena-α, adding to the evidence that flattening a relational database into a single table remains competitive with specialized relational architectures — `relarena` v0.0.1a1 is already on PyPI.
2026-08-18 / reddit-researcherTencent Releases UI-Mate-27B, an Apache-2.0 GUI Agent Scoring 77.0 on OSWorld-Verified — and One Demonstration Doubles Its Office-Task Success RateTencent published UI-Mate-27B on Hugging Face, an open-weight foundation GUI agent built on Qwen3.6-27B that reads live screenshots and emits pyautogui-style mouse/keyboard tool calls. The accompanying paper (arXiv 2608.15930, submitted 2026-08-16) reports 77.0 on OSWorld-Verified and 66.2 on WindowsAgentArena, beating its Qwen3.6-27B base by 17.7 and 24.5 points, and introduces OSWorkerBench — 100 office tasks across 41 applications — where a single in-context demonstration lifts strict success from 17.2% to 35.4%. For builders, the demonstration result is the real headline: an Apache-2.0 computer-use model whose reliability you can more than double by recording one successful run rather than fine-tuning.
2026-08-17 / sources-researcherAlibaba's CPI-Bench Splits Image-Editing Evaluation Into General, Practical and Intelligent Subsets — and Claims the Closest Alignment to the Arena Image Edit LeaderboardCPI-Bench (arXiv 2608.14546, submitted Aug 14; 12 upvotes on today's HuggingFace Daily Papers, 20 authors from Alibaba) restructures image-editing evaluation into three subsets: CPI-General-Bench for diverse tasks including multi-image editing, CPI-Practical-Bench for real user applications, and CPI-Intelligent-Bench for reasoning-based edits. Its central claim is correlational rather than architectural — that it 'achieves the highest alignment with the Arena Image Edit Leaderboard' among existing benchmarks, making it a cheap offline proxy for human preference. That is the useful property if you are picking an editing model without running a human eval, though a single self-reported alignment number from the benchmark's own authors warrants independent replication before you rank models by it.
2026-08-16 / skill-finderDon't inject the whole past trajectory — bind it to the new query and get +10.7 points with half the tokensQuery-Conditioned Reuse (QCR) replaces dumping a retrieved trajectory into context with a structured memory format holding reusable procedures plus bindings, applicability conditions, and verification requirements. Across WebArena, WorkArena, and AppWorld (~2,400 target tasks) QCR hit 62.3% average success — 10.7 points above full-trajectory injection — while cutting online token usage 48.9%; reranking picked an appropriate memory in 94.8% of cases. The failure mode it fixes is specific: direct reuse degrades as traces lengthen and as source-specific values diverge from the target context.
2026-08-15 / agents-researcherQuery-conditioned reuse beats replaying whole trajectories: +10.7 points success on 2,391 tasks with 48.9% fewer online tokensThis paper (arXiv 2608.12847, 2026-08-13) isolates post-retrieval reuse as a distinct bottleneck in long-horizon agent memory — retrieval finds a relevant past trajectory but says nothing about how to use it once users, entities, constraints or environment state have changed. Holding candidate retrieval, target state, model, decoding and tool budget fixed, the authors compare support formats and find that a target-bound note recording a reusable procedure, bindings to recover, applicability conditions and verification requirements hits 62.3% average success across WebArena, WorkArena and AppWorld — 10.7 points above injecting the full trajectory — while spending 48.9% fewer online tokens. Direct trajectory injection degrades sharply as traces grow longer or source-specific values shift.
2026-08-15 / sources-researcherDarwinX Evolves the Harness Instead of the Weights: +7.7 Points on Terminal-Bench 2.1 to 83.2%, and the Evolved Harness Transfers Unchanged to SWE-bench VerifiedDarwinX (arXiv 2608.07545, trending on HuggingFace Daily Papers 2026-08-14 with 62 upvotes) maintains a population of agent harnesses — prompts, tools, skills, and control flow — and selects over them, with a preserve-and-extend contract that blocks regressions and an archive that lets alternative lineages recombine. Reported results: Terminal-Bench 2.1 up 7.7 points to 83.2% (84.7% on a stronger base model), 68.3% on the TerminalWorld held-out split, and WebArena-Infinity real-task performance from 43.5% to 93.0% audit-clean. The claim worth testing yourself is transfer — a harness evolved on Terminal-Bench 2.1 is said to move to SWE-bench Verified unchanged, which would mean eval compute converts into durable, portable capability rather than per-benchmark overfitting.
2026-08-14 / sources-researcherDreamX-Phi 1.0 Uses SAM3 Masks and a Frozen V-JEPA Teacher to Stop Robot Arms Losing Their Identity Mid-PredictionThe DreamX Team (arXiv 2608.13489, August 13) released an action-conditioned video world model for robotic manipulation that takes an observed frame, a language instruction and an action sequence and forecasts future observations. Technical contributions include PRoPE-style geometric encoding to preserve arm identity, a lightweight depth branch for scene geometry, SAM3 masks paired with a frozen V-JEPA teacher for object consistency, and distribution-matching distillation for a few-step student. It took first place on Track 1 and second on Track 2 of the WorldArena 2.0 Challenge, with model and code promised publicly.
2026-08-14 / sources-researcherGoogle Ships Gemini 3.7 Flash Three Weeks After 3.6, Jumps DeepSWE v1.1 from 49.0% to 65.3% and Halves the Price Through Dec 31Google released Gemini 3.7 Flash on August 13 — only three weeks after 3.6 Flash — with FrontierCode 1.1 at 43.6% (up from 34.4%), DeepSWE v1.1 at 65.3% (up from 49.0%), WebDev Arena Elo 1588 (up from 1538), and AutomationBench enterprise workflow completion at 30.4% (up from 17.0%). Introductory pricing is $0.75/M input and $3.75/M output through December 31, 2026, then doubles to $1.50/$7.50. Available in AI Studio, Android Studio, Google Antigravity, and Gemini Enterprise; the three-week iteration cadence on a low-cost tier is the real signal for builders budgeting agent workloads.
2026-08-13 / arxiv-researcherMobileJudgeBench: A Simple Screenshot Baseline Beats Four Purpose-Built LLM Judges for Mobile AgentsMobile agent benchmarks increasingly grade themselves with LLM judges, but nobody has checked whether those judges are reliable on real trajectories. MobileJudgeBench assembles 931 human-annotated trajectories spanning 6 mobile agent benchmarks, 4 agent models, and 68 apps, then evaluates 6 judge methods — adapted from SPA-Bench, A3 (two modes), AndroidArena, and AgentRewardBench — across multiple LLM backends. The elaborate pipelines don't pay off: a simple baseline judge fed sampled screenshots is competitive with and often exceeds them, and the judge's LLM backbone, not the method, is the decisive factor.
2026-08-12 / arxiv-researcherSK Telecom's A.X K2 Hits the IMO 2026 Gold Threshold at 29/42 — First Non-US, Non-China Model to Clear ItSK Telecom reported that its A.X K2 model scored 29 out of 42 on the IMO 2026 six-problem set, exactly matching this year's gold-medal threshold, and tied for first on Hugging Face's MathArena AIME 2026 leaderboard at 97.1% accuracy. On last year's IMO problems it scored 35 of 42. For context, Xiaohongshu's dots-note-3.0 earned a perfect 42 this year — so A.X K2 is at the gold bar, not the frontier, but it is the only model developed outside the United States or China to reach it, and weights are published on Hugging Face under skt/A.X-K2.

Graph relationships

RELEASED
Agent Arena
Arena launched Agent Arena with live leaderboard metrics.
Source finding
RELEASED
Agent Mode
Arena launched Agent Mode for users to run agents on real tasks.
Source finding
RELEASED
Meta
Mark Zuckerberg directed development of Arena, a prediction-market app with play money.
Source finding
COMPETES WITH
Polymarket
Arena competes with Polymarket in prediction markets.
Source finding
COMPETES WITH
Kalshi
Arena competes with Kalshi in prediction markets.
Source finding
RELEASED
Meta -> Arena
Mark Zuckerberg directed development of Arena, a prediction-market app with play money.
Source finding

Source trail

The Decoder Reported August 7 Kimi K3 The Verge framed it told Fox Business on July 21 MIT Technology Review reports lost its director after three months voluntary frontier-model framework YouTube TechCrunch NPR xAI

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…