Research hub
Curated reports for the two-circuit GRPO work: current weekly status first, then the evidence chain behind anchor projection, dataset transfer, and stability at longer coordination intervals.
Most important tab
The current readout for people who only have time for one page.
The stale side can predict future weights well enough on the easier GSM8K trace, but the same projector collapses to do-nothing on harder Math traces. The practical answer is a coherence gate: project only when consecutive updates are aligned.
Curated evidence
Only the reports that carry the current decision chain are copied into this repo.
Decomposes the two tasks. The rank-1 spatial structure only dents (EVR₁ ~90-96% on both), but forward linear projectability collapses (op ratio 0.90 → 1.00). Reconciled by two timescales: the cumulative direction is stable on both, while raw per-step updates align only on GSM8K (0.86 vs 0.15). Rank-1 is necessary but not sufficient for projection.
Re-runs the rank-1 analysis on Big-Math (EXP-58). Rank-1 structure replicates, but the single line's durability degrades on the harder task: ~6× more of the move lands off the line (25% vs 4.5%) and the direction rotates further over the run (|cos| 0.46 vs 0.64).
Seven communication-efficient runs varying the merger's correction weight α and anchor-memory β, plus a dense control. α=0.25 with a fresh anchor (β=0.00) is the stable winner; only the stalest anchor (β=0.90) collapses. Dense stays the ceiling.
Live training on Big-Math: no compressed variant beats dense, and plain stale signed EMA every step (no projection) beats every projection variant. A GSM8K appendix shows the coherent-trajectory contrast that explains why.
Integration rollup for the projector lanes. Fixed damped-linear wins on GSM8K, but is not adopted as a universal default because Big-Math collapses to do-nothing.
EXP-61 confirms the harder Math trace is only marginally projectable at the freshest anchor and do-nothing-optimal at operational staleness.
RELEX-style rank-1 structure replicates, but the paper form is the wrong objective for short-horizon staleness repair. Use rank-1 as position state, not blind extrapolation.
Live training result for communication-efficient pipeline-parallel GRPO at longer cadence, useful as the bridge from offline projection geometry to end-to-end stability.
One-page rollup
The minimum context needed to understand where the reports came from.
EXP-43, EXP-44, EXP-45, and EXP-58 establish the fp32/bf16 weight trajectories, sweep engine, scorecard contract, and R2-verified checkpoint collection.
EXP-47 finds fixed damped-linear projection at ratio 0.9396 with h_safe around 30 global steps; EXP-49 adaptive gains are real but sub-threshold.
EXP-60 and EXP-61 show consecutive update cosine near 0.15, so operational projection collapses to hold-stale except at the freshest anchor.
| Date | Name | Takeaway |
|---|---|---|
| 2026-07-08 | Signed-EMA α,β sweep | Moderate correction + fresh anchor is stable at 20/20; only the stalest anchor (β=0.90) collapses. |
| 2026-07-06 | Weekly update | Coherence-gated projection is the current design direction. |
| 2026-07-05 | Math-only ablation | Fresh-anchor projection clears about 1%; operational Math remains do-nothing-optimal. |
| 2026-07-04 | MOAT projector verdict | GSM8K win does not generalize; no universal projector default. |
| 2026-07-03 | Rank-1 trajectory analysis | Rank-1 state is useful for long-range position, not short-range displacement prediction. |
Older report set
These findings matter, but the duplicated full pages were removed from the publishing repo.
The same signed EMA merger is near the dense ceiling at K=5, but falls below the no-merger floor by step 100 at K=20. The diagnosis is stale/off-policy anchor gradients, with cadence as a secondary contributor.
The compressed trainer and rollout engine disagree too strongly for plain truncated importance sampling to cleanly absorb; this remains a separate correction path from anchor staleness.