Skip to content

SOTA sweep 2026-10-07: Cross-Tokenizer OPD, TRACE/TRIAGE, NeMo-DCR, NP-OPD, MASKerade - #40

Merged
cursor[bot] merged 1 commit into
mainfrom
cursor/sota-sweep-2026-10-07-17b6
Oct 7, 2026
Merged

cursor[bot] merged 1 commit into
mainfrom
cursor/sota-sweep-2026-10-07-17b6

Conversation

@jancompute

Copy link
Copy Markdown
Owner

Weekday Librarian sweep after PR 39 (arXiv watermark was 2610.06833; main ab05f6d). Fifteen active/niche methods, four new first-hop tasks, two paper-only notes. Default placement is active beside incumbents (sota_for: []). No locked current_sota retarget (CISPO, VISTA, OPD, Muon2, AdamW, Miles, Quartet-II, ES-reasoning, Category-Aware SWE Experts, SAPO, DeepSeek-V4, Kimi-K3, CanvasAnneal, DiffusionOPSD). No supersession edges.

python3 -m library index / export / validate: 0 errors. dist/graph.json is regenerated in this PR (1134 nodes, 4812 edges; was 1083). Every new node id is present in dist/graph.json.

What a training agent should care about

  • Cross-tokenizer OPD: do not expand leftover-token alignment with span MSE. Strict 1:1 already covers most student tokens; top-16 shared-vocab reverse-KL matches full shared-vocab OPD.
  • FP4 RL numerics: TRACE caches rollout mantissa/scale into the train grid; TRIAGE rebalances amplifying negative-advantage/negative-gap updates. Complementary, no H2H; neither is Quartet-II.
  • 1T disaggregated RL: NeMo-DCR bit-exact delta refit (~1% of BF16 weights/step) is the weight-sync plug-in beside Miles, not a new engine.
  • Low teacher/student overlap: NP-OPD adds negative-policy rollouts to OPD (related to NSD, not a replacement).
  • Adam shared-β: 200-update pilot + cubic rule; 40.7% lower mean relative val gap vs β=0.95. Beside AdamW, not a Muon2 retarget.
  • Muon spectral maps: BulkBoost argues two-band MP reweight is enough; keep Muon2.
  • MoE load-balancing: DRMoET is a drop-in DRO objective on FLAME-MoE; keep V4/K3 as architecture.
  • Agentic MoE RL: constrain routing to operation types beside ESRL/RPB; SAPO stays the MoE loss.
  • SWE recovery: FC-SWE reuses failed patch + verifier feedback; Category-Aware SWE Experts stay first hop.
  • Sparse-reward scaffolding: RGPO (adaptive GT rationales) + GA-GRPO (optimal guidance weight theory). CISPO stays Pass@1.
  • OPSD drift / OPD-for-safety: privileged-context content drives drift more than source; a backdoored safety teacher transfers hidden behavior under OPD (3% poison → up to 70% ASR).
  • dLLM recycle / dense-to-MoE: A2D adds an AR post-train delta to a converted dLLM; MASKerade learns binary-mask experts over a frozen FFN.

Ingested methods

Method arXiv Why (paper metrics) Neighbors Code
method:cross-tokenizer-opd 2610.08448 Strict 1:1 covers 85.57–96.98% of student tokens; k=16 retains ≥96% of full shared-vocab OPD. Top HF Daily 10-07. opd none
method:trace 2610.07767 Up to 5.4× rollout; joint FP4 W/A + FP4 KV vs BF16-level RL quartet-ii / triage none
method:triage 2610.07043 Up to 2.3× rollout vs BF16 on native NVFP4; segment-level rebalance quartet-ii / trace none
method:nemo-dcr 2610.08430 ~1% of BF16 weights change/step; 1T relay-tree 150s vs 87.5 min miles / rufus-air NVIDIA-NeMo/RL PR #2444 released
method:np-opd 2610.07874 Negative-policy rollouts when teacher/student overlap is low opd / nsd naver-ai/np-opd released
method:adam-beta-cubic 2610.08624 40.7% lower mean relative val gap vs β=0.95 on 11 workloads adamw-optimizer / muon2 AlbertoFdezHdez/Adam_beta_rule_cubic released
method:bulkboost 2610.07497 Two-band MP reweight 0.073–0.147% loss drop vs Muon flat (Freon 0.022%) muon2 / musec / orca none
method:drmoet 2610.07207 FLAME-MoE 10.3B seven-task avg 0.6767 vs 0.6625 / 0.6431 (NeurIPS 2026) deepseek-v4 / kimi-k3 MAPS-research/DRMoET released
method:hierarchical-moe-routing-control 2610.07332 >10-point success on evaluated agentic benches; operation-type experts esrl / rpb / sapo none
method:fc-swe 2610.07898 SWE-bench Verified 500 Resolved@1 41.7 vs GRPO 38.9 category-aware-swe-experts none
method:rgpo 2610.07342 Adaptive GT rationale scaffolding (NeurIPS 2026) cispo / mintrl VietHoang1512/rgpo released
method:privileged-context-drift 2610.07842 Content drives KL 5.1× more than source. Niche evidence, not a trainer. vista / u-opsd / opsd-collapse-review none
method:a2d 2610.08108 AR post-train delta on converted dLLM approaches direct diffusion PT and composes with it diffusion-opsd / canvasanneal none
method:maskerade 2610.07809 Four 2:4 mask experts, top-2, frozen FFN deepseek-v4 / ce-moe Ming-K9/MASKerade released
method:dart-es 2610.06993 GSM8K 73.53 vs ES 72.07 vs GRPO 73.26 es-reasoning / cispo szs777/DART-ES-Code released

Claim notes (no new method)

Paper arXiv Placement Code
paper:ga-grpo 2610.06861 Theory/evidence on RGPO / MInTRL / CISPO. λ*=σ0²/(σ0²+Rmax²δ²T); 31% fewer GPU-h. Does not retarget CISPO. none
paper:opd-safety-backdoor 2610.07654 Warning on OPD used for safety alignment. 3% poison → up to 70% ASR; more epochs amplify. Does not retarget OPD matching. none

New tasks

  • task:cross-tokenizer-opd — first hop method:cross-tokenizer-opd. Reverse redirect from task:student-distillation.
  • task:fp4-rl-train-rollout-alignment — dual-active TRACE + TRIAGE. Reverse redirect from task:fp4-hardware-training. No H2H, so both listed; methods keep sota_for: [].
  • task:diffusion-lm-ar-delta-recycle — first hop method:a2d. Reverse redirect from task:posttrain-diffusion (CanvasAnneal stays the discrete-DLM canvas niche there).
  • task:dense-to-moe-upcycling — first hop method:maskerade. Reverse redirect from task:pretrain-moe-frontier.

Placement decisions changed from the prompt

  • Hierarchical routing: primary home is task:math-code-rl-moe beside ESRL/RPB (as suggested), with a mention on task:outcome-only-long-horizon-agent-rl rather than making it a CANOPY competitor. SAPO / CANOPY unchanged.
  • A2D: new task:diffusion-lm-ar-delta-recycle rather than parking it as a first hop on task:posttrain-diffusion (that task remains image/flow reward alignment; CanvasAnneal stays the discrete-DLM canvas niche).
  • TRACE vs TRIAGE: dual-active on the new FP4 RL task; no supersession (no head-to-head).
  • Privileged-context drift: status: niche method+recipe (evidence), not active trainer.
  • GA-GRPO and OPD-safety-backdoor: paper-only; no method nodes.

current_sota changes

None retargeted. The only new current_sota listings are on the four new tasks (methods still sota_for: []).

Skipped

None of the listed candidates were already in the graph. None were dropped. Freon/Spectra (BulkBoost baselines) were not created as extra nodes.

Open in Web Open in Cursor 

…, NeMo-DCR, NP-OPD, MASKerade)

Add-on plug-ins beside incumbents. No locked current_sota retarget. dist/graph.json regenerated (1134 nodes).

Co-authored-by: Jack Large <j8ckfi@users.noreply.github.com>
@cursor
cursor Bot merged commit 8a4bb50 into main Oct 7, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants