2026-09-27

From sources/papers-daily/hf-daily-2026-09-27.md — 25 entries, 17 new after dedup against pages and yesterday's one-off mentions. Upvote counts are that community's popularity signal and nothing more.

SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue — arXiv 2609.26780 (1.65 — the day's highest score)

• TL;DR: general agent memory retrieves content and loses who said it about whom. Two tracks — speaker-labelled verbatim messages and derived state in person-level and group-level views — joined at query time by entity, event and time. 47.9 / 69.2 / 61.9% binary accuracy on GroupMemBench / SocialMemBench / EverMemBench, 70.85% on all 1,986 LoCoMo questions
• Why read it: it is the counter-example to the week's run of papers deleting components. Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents argued curation should wait for the query; this argues speaker attribution must be built at write time because it cannot be recovered from an interleaved transcript afterwards. The two are the same disagreement stated precisely, published a day apart, neither citing the other. Its own headline number is the ablation: RL on the writer alone moves 57.38 → 68.20, +10.82, architecture held fixed
• → SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue

Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms — arXiv 2609.27321 (1.65)

• TL;DR: pipelines "construct an environment before defining its outcome rule", then align the two afterwards. VHD-Play solves a mathematical model first and has a setter render its decision process as stateful tools, so executable dynamics and trajectory scoring are inherited from the same solved model. 3,300 environments at a few cents each; Qwen3.6-35B-A3B 0.204 → 0.815 mean agentic score, transfer to eight unseen mechanism families and to function calling, travel planning and e-commerce; on E-Commerce Bench it completes every run without bankruptcy and exceeds Qwen3.7-Max
• Why read it: the ablation, not the 4× headline. Comparing written-out problems against stateful versions, most of the learnable gap turns out to be stateful interaction rather than the underlying problem solving — the most direct evidence this wiki holds for why agentic and reasoning benchmarks come apart. The headline itself is an undefined scale: nothing read says what the agentic score measures or where its ceiling is
• → Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms

ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds — arXiv 2609.30199 (1.49)

• TL;DR: two sandboxes — AlienCode (31 discovery targets, 70 tasks) and AlienLogic (24, 70) — whose rules are executable, so answers check exactly, and which conflict with familiar knowledge, so recall alone cannot solve them. Each ships a flawed manual. Across 10 systems the strongest do acquire and apply unfamiliar rules, but performance varies substantially across trajectories and continued exploration can stall or reverse earlier gains
• Why read it: the contamination control is the contribution — most agent benchmarks' defence against pre-training leakage is hoping the task is new, and making the ground truth deliberately alien is a control. Pair it with Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms from the same snapshot: the same scarce resource — a verifiable unfamiliar mechanism — manufactured once to train on and built once to test with, a day apart, neither citing the other. No score is reported, so the third finding is the one to carry: an agent that gets worse the longer it explores fails differently from one that plateaus
• → ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds

• The snapshot intake reads for figures, not for names. Today's +52-day capture was in this repo's own weekly leaderboard files for 13 consecutive captures, and InclusionAI has been in every Artificial Analysis snapshot here since 2026-07-30. Yesterday's finding was "the lab is in no tier of sources.yaml" — true, and not the whole failure. Nothing sweeps the Creator column for labs the wiki does not have. Carried to the W39 lint
• A Sunday snapshot is missing and nothing failed. sources/evals/lmarena-2026-09-27.md was due at 07:00 KST; commit a46cb05 carries artificial-analysis-2026-09-27.md and hf-daily-2026-09-27.md and no LMArena file. Last capture is 2026-09-20. The scrapers were not run from here, per standing policy. Carried to the W39 lint as check 2o
• This wiki asserted an absence that was not there. World Models, created 2026-09-26, said there was no instrument on which the generative and predictive senses of the term could be compared. HappyWorld-Bench is dated 2026-09-21 — six capabilities over video / spatial / embodied tracks, 1,138 prompts / 300 scenes / 254 cases, 31 systems, human A/B Elo alongside behavioural metrics, spatial models at at best 70.14% placement and 73.33% edit execution. Recorded on the page as a reading failure, not a field gap
• A first-party index listed different items on consecutive reads. www.anthropic.com/news gave seven September items today where yesterday's run recorded four from the same page. Six are held; the seventh, "The Situation Report" (09-22, Features), is not — see 🔇. An index whose contents vary between reads is a dedup hazard, and nothing in this pipeline would report it
• Gemini 4's timing, held at arm's length. Koray Kavukcuoglu is reported (The Information, via one pass, 09-23/24) to hope to launch Gemini 4 "much earlier" than the end of 2026, with no month named. Single-pass, second-hand and undated, so it is not written into Gemini 4's Release Date — the same discipline that made yesterday's avatar hold the right call

Eight pages, no new entity or concept page — today's captures landed on entities and concepts that already existed, which is the first run in four days with nothing needing your review on that front.

• Ling-3.0-tiny (new) — MIT, +52 days, and the only page here whose checked figures come from this repo's own captured leaderboard rather than a summary
• Muse Realtime Avatar (new) — Released reads not yet; four of seven Spec rows are unknown, including License, which is not written as proprietary despite every sibling being closed — a licence carried over from a sibling is a guess
• SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue (new) · Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms (new) · ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds (new) · HappyWorld-Bench (new) · Rufus-Air: An Open LLM Post-Training Recipe (new) · Hunyuan-A13B Technical Report (new)
• A page deliberately not created: Hunyuan-A13B has no model page. The snapshot dates its technical report 2026-09-23; every search pass dates the weights 2025-06-27. A model page must assert a Released value, and fifteen months apart is not a rounding difference. The material is on Hunyuan-A13B Technical Report and Tencent instead

Still needs your decision, unchanged from 09-26 — sources.yaml is manual and this run did not edit it. Yesterday's three requests stand (Runway; inclusionAI / AntLing as a sixth Chinese-rotation slot; a threshold exemption for the Transparency Coalition). Today adds a fourth that is not a source at all: a sweep of the Creator column in each Artificial Analysis capture against the wiki's entity list. It is one comparison over a file this repo already commits, and it would have surfaced inclusionAI on 2026-07-30.

• Ant Group (inclusionAI / AntLing): the Ling and Ring lines, five sibling leaderboard rows, and ## Strategic Position off _TBD_ for the first time — reporting a complication rather than a thesis, because a lab with a 1T model is not the specialist its two captured releases suggest
• Meta AI: the avatar adopted into Recent Activity; the ## Conflicting Reports hold rewritten to record its own resolution, including that its factual guess was wrong. The Private Processing on glasses half is still single-pass and still not adopted — what corroborates is distribution, not on-device privacy, and they are not the same claim
• Tencent: first entry since 2026-09-01; an 80B/13B open MoE with five domains claimed and zero scores, absent from the 269-row table this repo captures
• Z.ai: a third party's eight-stage recipe on GLM-4.5-Air-Base claiming to beat Z.ai's own post-trained release — with no benchmark figure, on a base two minor generations behind what the leaderboards list
• World Models: the missing instrument, found, dated five days before the page that said it was missing
• Agents (LLM Agents): SpeakerMem-R1 as the week's counter-example, plus two independent treatments of search agents in one snapshot — IterSynth's architecture and Rufus-Air's dedicated training stage
• Agentic Reinforcement Learning: VHD-Play removing the step where a reward gets written by hand, against PACT: From Credit Assignment to Critic Alignment proving what a correct one is
• Content Provenance (AI output marking): the three-model provenance table, and a product detail with a consequence — a cartoon default and a photograph input inside one feature, treated identically by one watermark
• R&D Automation Index: ExplorationBench and WhatWorkedBench, two benchmarks for the step this page's 26% figure assumes. WhatWorkedBench's result is the harder one: a Gaussian process fitted to the agent's own observations recovers effects better than the agent does — 0.632 → 0.698
• Open-Weights Policy Fight: "open weights" used this month for MIT and for a territory-excluded custom licence — a restriction type this page had not held
• Post-Training Scaling: the first fully documented recipe on the page, publishing no numbers
• index.md: 8 new lines, the Ant Group (inclusionAI / AntLing) line rewritten