Reward the writes that mattered
The Hy-Memory method, read in full: chunk-level GRPO with outcome credit attributed to the exact evidence chunks, trap probes for over-storage, and reward data built from failure surfaces — and no ablations yet to prove the attribution itself.
The owed Hy-Memory paper read discharges Issue 14’s open question — “attributed” means credit attribution to evidence chunks, not provenance — and reveals a 164,576-probe failure-surface reward set with trap probes and a three-hop examination, while deferring every ablation; Tencent also ships Hy-Nomas (persistent-KV interactive video worlds on one RTX 4090) and WeVisOne-VL (unified object+pixel, Apache-2.0, no benchmarks yet). Issue 16, coverage window 22–25 October 2026.
Week of October 22–25, 2026 · Issue 16
The Week in One Sentence
The Hy-Memory method read turned memory-writing into a credit-assignment problem whose central claim is asserted by design and unproven by ablation, Sarvam’s sovereign rate card went generally available with every planned-against price held, Real-SWE re-scored silently again — re-crowning Opus 5.5 and converging with the independent index for the first time — Open-SWE v2.0 finally shipped the harness trajectories promised since September, and Japan’s labs shipped eval-tooling velocity and a pre-release nudge but nothing above the bar.
The Big Story: the Hy-Memory method — reward the writes that mattered
What happened. (China; labelled owed read — the paper behind Issue 14’s release, queued then, discharged now.) We read arXiv:2610.22196 in full (Tencent Hunyuan 2026), and the first thing it settles is this desk’s own open question: “Attributed” Memory Reinforcement Learning means credit attribution of outcome reward to specific evidence chunks — not provenance, not citation. When a downstream question gets answered, the judge reports which chunk’s gold facts the answer recovered (c_q[t] ∈ [0,1]), and reward flows to the writes that supplied that evidence — not broadcast across the whole trajectory.
Under the hood — three ideas worth stealing. First, the framing: memory writing is “a policy for changing an external state under a long-horizon downstream objective” — not summarization; a writer must filter transients, preserve exact names and numbers, keep old and new state on transitions, and reconcile with the bank. Second, the training pipeline: on-policy distillation where the frozen flagship teacher scores the student’s own trajectories through the unmodified harness write path (no ideal demonstrations needed), then chunk-level GRPO — eight rollouts per window from independent bank copies, advantages normalized per chunk-position, so one rollout can be credited at chunk 2 and penalized at chunk 4. Third, the reward data: built from failure surfaces — omission, semantic inversion, stale-state confusion, secondary-fact loss, over-storage — yielding 164,576 recall probes, 8,599 trap probes (where the correct answer is “bank contains no such record”), and ~60.8k outcome questions with 222,395 gold-fact references. Evaluation is a three-hop examination: writer, answer model, and judge with strict information separation — gold facts visible only to the judge, rewards computed through the full deployable pipeline including retrieval over the written bank.
What the evidence says (vendor-run, with the counter-evidence at equal billing). First on all four harnesses; macro 55.49 vs Claude Opus 5’s 54.62 — but the authors themselves label the 0.01–0.02 margins on two harnesses as parity, and the number that actually matters is out-of-distribution: on held-out EverMemOS, +2.86 over Opus 5, driven by Travel SPS (+10.65) and Code Memory Bench (+13.66), with base→trained at +7.84. Then the gap this desk must carry: there are no ablations — the paper explicitly defers all of them, including the load-bearing one (attributed vs broadcast reward). With 20% subsamples and no repeated seeds, the central claim — that attribution matters — is asserted by design, not demonstrated by removal. The limitations section, to its credit, says so.
Why this matters. This is the most complete recipe yet for training agents that write to external state — and the transferable parts (failure-surface reward construction, trap probes, trajectory-scoring distillation) apply far beyond memory: any agent whose output mutates a persistent store has the same credit problem. The field’s honest frontier is right where this paper stops: attribution-shaped rewards are intuitive, and unproven. The follow-up paper that ablates them is the one to watch.
What I would test. The ablation the paper skipped, approximated from the released artifacts: train (or fine-tune) two writers on a small corpus — one with attributed credit, one with broadcast — and diff the banks: which writes survived, which polluted, and whether the OOD scenario gains (Travel, CMB) survive. Cheaper version: run the released writer against Opus 5 on your own traffic and score pollution, the metric none of these tables carry.
Elsewhere in AI Engineering
Hy-Nomas: interactive video worlds, persistent KV, one 4090 (China) — Worth Testing
What happened. (China.) Tencent’s Hy-Nomas (~Oct 24/25) generates video worlds segment by segment under free-form multi-turn instructions and camera control, with a persistent KV cache carrying context between segments — 720×1280 at 24 FPS, 8 distilled denoising steps. The architecture is a segment-autoregressive diffusion transformer (54 double-stream MMDiT blocks, 3D RoPE) with a Qwen2.5-VL-7B text encoder and camera extrinsics via cross-attention. Apache-2.0, with code, a tech-report PDF, a web demo, a robotics LoRA, and Hy-Nomas-Bench (280 cases across multi-turn text, camera navigation, interleaved control, and spatial memory). The deployment line is the headline for engineers: a single-RTX-4090 config (INT8 DiT, INT8 KV, FP8 decoder, CPU offload).
Why this matters. Two ledger threads converge: memory continuity (the persistent cross-segment KV cache is the same state-across-time problem this desk has tracked from Engram to Hy-Memory — now inside a generative loop) and compression (a 24 FPS interactive world model on a consumer card is the quantization thesis at its most concrete). The card publishes no scores — the bench suite ships for that; first numbers are the watch.
GA day on the Indian rate card (India) — Worth Reading
What happened. (India. Labelled update on Issues 6 and 12 coverage.) October 22, per the docs changelog: GLM-5.3, Gemma 4 31B, DeepSeek V4.1 Flash, and the Responses API (/v2/responses) became generally available — “no beta whitelisting is required for these endpoints or models.” The pricing page, re-read today, shows all four model rows flipped from Beta to Available with prices unchanged from the October 10 read: sarvam-105b ₹15/₹5/₹60, gemma4 ₹35/₹10/₹90, glm5.3 ₹126/₹20/₹390, deepseekv4.1-flash ₹25/₹0.50/₹105 (₹ per 1M, GST-exclusive; reasoning tokens billed as output).
Why this matters. The arc this desk has tracked since September closes clean: glm5.3 arrived beta on a rupee rate card (Issue 6), took a quiet cut (Issue 12), and now graduates with pricing held — the numbers readers planned against survived. The strategic line completes too: an Indian platform now GA-serves Chinese and US open-weight flagships alongside its own 105B — distribution as the sovereign-stack product. One gap, labeled: the Responses API has no pricing section on the rate card; whether it bills identically to chat completions is unverifiable from the page.
The Bulbul deadline narrows (India) — Worth Watching
What happened. (India. Labelled update on Issue 15’s printed item.) The changelog’s EOL row, re-read this cycle, is narrower than our last print: Bulbul V3 “reaches end of life on Voice Agents on November 10, 2026” — and “Bulbul V3 remains on the Text-to-Speech API.” Migration for Voice-Agent users runs through Settings → Speaking; sixteen days remain at this issue’s date.
Why this matters. Scope is action: teams using V3 through the TTS API have no deadline; teams with V3-voiced agents have sixteen days and a settings-screen path. This deadline was re-entered in Issue 15 after a cap cut erased it, and the narrowed scope is the second precision pass — a blanket “V3 retires November 10” reading would have sent API users migrating for nothing.
Go deeper: changelog
The measurement layer moved twice: a re-crown, a debut, and the promised harnesses (US) — Worth Reading
What happened. (Labelled update on Issue 14’s coverage.) Real-SWE 1.1 re-scored again inside the same version — now 17 models — and the top flipped back: Claude Opus 5.5 jumped 42.40 → 56.80% to take #1 (GPT-6 Astra rose to 56.00), while Claude Sonnet 5.5 debuted at #3 — 51.20% at $10.09 per rollout — and the two rows this publication printed in Issue 14 sat frozen (GPT-6.1 Sol 47.20, Fable 5.1 45.60). A partial re-run, then, with no revision note, no paper, and no explanation — the third unexplained move on this instrument in five weeks; every delta is our arithmetic over two dated snapshots. And the data layer moved with it: Open-SWE-Traces shipped v2.0 — new rollouts powered by Qwen3.8-Flash-Next (xhigh reasoning) across five harnesses, including the Claude-Code and OpenCode trajectories promised since September — closing this desk’s twelve-cycle watch (data landed October 23 per the API; the card’s “[10/26]” date line repeats its forward-dating quirk; ~438 GB, CC BY 4.0; low/medium reasoning efforts “coming soon”).
Why it matters. This publication printed “Real-SWE crowns GPT-6.1 Sol” three weeks ago; the instrument has since re-scored twice without saying why — and the buyer’s rule that follows is the one China’s items keep writing: instruments move, so date-stamp every number you act on. Two things genuinely improved, though. The top of Real-SWE now agrees with AA’s index (Opus 5.5 led AA at 58 on this desk’s October 14 read) — the first top-of-table convergence after three cycles of disagreement. And Open-SWE’s harness coverage finally includes the two most-widely-used commercial-agent traces, which makes the dataset’s exploit-audit methodology testable against Claude Code itself — the check this desk flagged in Issue 5. Note the cross-region texture: the training data runs on Qwen3.8-Flash-Next, the weights line China’s ledger has tracked since Issue 3.
Go deeper: Real-SWE 1.1 · Open-SWE-Traces
No Japan items this issue: the window’s Japan-lab output was an eval-suite maintenance burst on NII’s simple-evals-mm (ZeroBench, MMMU-Pro, AIME 2026, and BabyVision added within 36 hours — while the org’s own public leaderboard stays frozen with no new model scored), the second pre-release nudge on the full-duplex llm-jp-moshi line, and chores. Stated rather than padded, per the no-filler rule; the release watch continues, and the eval-tooling velocity is ledgered.
Research Radar
The window’s research story leads the issue as the Big Story — the Hy-Memory method paper, read in full (paper-notes in the packet). No in-window China paper met the read-beyond-abstract bar; Hy-Nomas’s tech report PDF (in-repo, unread this cycle) and the WeVisOne-VL paper-to-come are queued.
Repo Radar
Tencent-Hunyuan/Hy-Nomas (China)
- What’s actually in it: the world-generation model (Apache-2.0), inference configs including the single-4090 path, Hy-Nomas-Bench (280 cases, four tasks), the tech report PDF, and a robotics LoRA adapter. Inspection depth: card in full; code and PDF not opened this cycle.
- What you can run: interactive multi-turn video generation with camera control — on one consumer GPU, tonight.
- What’s missing: parameter count, any numeric benchmark results, and (from the wider ecosystem) any third-party Ascend or non-NVIDIA numbers.
- Why open it: the cheapest interactive world model we’ve ledgered, and the cleanest test of whether persistent-KV world consistency holds beyond vendor demos.
Tencent/WeVisOne-VL (China)
- What’s actually in it: the unified object+pixel model (~4.83B F32) and 2B sibling, pixel-decoder checkpoint, code, two demo Spaces. Inspection depth: card; code not executed.
- What you can run: detection, pointing, counting, OCR localization, segmentation, and alpha matting from one model.
- What’s missing: benchmarks, paper, video support — everything that would let you compare it to the grounding field.
- Why open it: the token-unification architecture is worth reading even before the numbers arrive; Apache-2.0 makes experimenting free.
nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 (US)
- What’s actually in it: (Owed read — GA October 16, held in this desk’s ledger since the October 18 cycle.) A ~4B “looped (latent-recurrent)” nGPT model: Input / shared Recurrent / Output layer blocks, and the headline property — one checkpoint, three compute budgets: set the loop count to 1, 2, or 4 at inference, no retraining (two loops is the default). Last-loop KV caching keeps KV memory at one-loop size regardless of loop count. 256K context; pre-trained natively in NVFP4 for ~4.5T of its 6T tokens; SFT + multi-environment GRPO after. Inspection depth: card in full; not executed. OpenMDW-1.1 (card field).
- What you can run: the elastic-compute ladder on Hopper/Blackwell, plus MXFP8 (~1.7–1.8× faster than BF16) and NVFP4 (~2.2–2.4×) variants — and a GGUF Q5_K_M (~3.67 GB) for llama.cpp, putting looped-latent inference on consumer hardware.
- What’s missing: English only; Hopper/Blackwell only for the accelerated paths; and the vendor table’s loop-ladder (SWE-bench Verified 50.3 → 59.5 → 61.8 across 1/2/4 loops; Terminal-Bench 2.1 12.5 → 18.1 → 22.2) is self-reported, like every number this issue carries.
- Why open it: compute elasticity from a single checkpoint is the deployment-side twin of this issue’s memory threads — pay-per-thinking, chosen at request time; the loop ladder is the cleanest published demonstration of it.
My Read on the Week
(Synthesis and interpretation — not reported fact. Written after a cross-item pass spanning all desks’ items; Japan published none — see the note closing Elsewhere.)
1. Credit assignment is becoming the training question for state-writing agents. Hy-Memory attributes outcome reward to the exact chunks that earned it; LongCat’s config study (Issue 12) attributed quality to the harness over the model; Draft-KV’s audit (Issue 9) found gains hiding in interfaces. The same instrument — who actually caused this — keeps deciding what’s real across the quarter’s papers, and Hy-Memory’s gap is the instructive one: the field builds attribution-shaped rewards faster than it ablates them. The un-ablated method plus this issue’s experiment (run the ablation at notebook scale) is the pattern this publication should keep repeating: when the paper skips the control, name it and hand readers the recipe.
2. State and compute both went elastic in one window. Hy-Nomas keeps a KV cache alive across generated video segments on one consumer GPU; Hy-Memory trains a writer whose whole job is curating persistent external state; Dory runs one checkpoint at three compute budgets with last-loop KV holding memory flat. Whether the state is a memory bank, a cross-segment cache, or a recurrent loop’s KV, the design direction is identical — make persistence a first-class, resizable quantity, decoupled from the compute that reads it. Watch code agents adopt it next: session-persistent banks are already halfway there.
3. The numbers economy is the story behind every item. Artifacts shipping without numbers (WeVisOne-VL’s grounding tables, Hy-Nomas’s bench, DeepInfer’s sweeps — 21.9k stars now building against a frozen main); instruments moving without notes (Real-SWE’s third unexplained re-score in five weeks, now converging with AA’s index for the first time); the data layer quietly crossing regions (Open-SWE v2.0’s trajectories generated on Qwen3.8-Flash-Next); and even the eval-battery layer documenting its own noise floor (the Indic suite’s methodology note: LLM translation, automatic checks, partial human review, licenses per-dataset). Numbers arrive late, move silently, or don’t exist — and the reader’s defenses are the ones this issue keeps arming: date-stamp every figure, demand the instrument before the ranking, and treat a missing table as an invitation, never a claim.
What I’d Watch Next
- The season’s missing tables — the Hy-Memory ablation paper (attribution vs broadcast, the method’s open flank), first numbers for Hy-Nomas’s bench, WeVisOne-VL’s grounding tables, and DeepInfer’s sweeps: four community-runnable paths to the numbers the releases didn’t carry.
- Real-SWE’s next move and Open-SWE’s low/medium efforts — this publication has printed two crowns from an instrument that re-scores without notes; the next move either explains itself or forfeits the citation. The dataset’s remaining reasoning-effort tiers land next.
- SAS’s would-be reproducer — issue #2’s author gets a GPU-time answer or gives up; the license question (eleven checks, still absent) becomes urgent the moment training starts.
One Thing I’d Experiment With This Week
Ablate the attribution — the experiment the paper deferred, run at notebook scale. You don’t need 128 H20s to test whether attributed credit beats broadcast credit; you need a small bank, two training runs, and honest scoring.
- Setup: take the released Hy-Memory artifacts (or your own writer on any framework with a memory bank). Build a miniature reward set from your failure surfaces — 50 omitted facts, 50 stale-state cases, 50 traps (“bank contains no such record”) from real traffic.
- Conditions: two fine-tuning arms on the same base model — one with per-chunk attributed reward (judge marks which evidence chunk each answer used), one with the same total reward broadcast evenly across the window’s writes. Same data, same steps, same bank config.
- Measure: bank composition per arm (durable-fact retention, transition correctness, pollution rate — stale/wrong entries per hundred writes), downstream answer accuracy, and bank size. Decision output: the ablation table the paper didn’t ship, on your own domain — and if broadcast matches attributed at small scale, you’ve learned the attribution machinery matters only where trajectories get long, which is itself the finding worth knowing before you adopt either.
All vendor-reported figures are labeled as such; if independent evidence doesn’t exist yet, that’s stated rather than papered over. This issue’s desk ledgers live under research/2026-10-25/.
References
Citation
@online{dimsum_editorial2026,
author = {dimsum editorial, the},
title = {Reward the Writes That Mattered},
date = {2026-10-25},
url = {https://news.llm101.in/issues/2026-10-25/},
langid = {en}
}