Research hub
Communication-efficient RLVR
The two-circuit GRPO research hub: weekly status, curated reports, and the full run record.
Most important tab
Weekly Update
The current readout for people who only have time for one page.
Where the recipe holds, where it breaks, and where the gap is born
The current recipe works at the scale it was built on. At larger models, and more importantly at longer reasoning generations and agentic-style tasks, training collapses late even after large early gains. The main cause is now identified as the gap between the dense sampler and the compressed trainer, and it is being actively worked: pricing each pipeline cut shows the gap rises with cut depth, and a shallow-compressed layout cuts the starting gap by 28 percent.
Curated evidence
Essential Reports
Only the reports that carry the current decision chain are copied into this repo.
Where the gap is born: pricing each pipeline cut
Solo probes that compress one pipeline cut at a time show the sampler-trainer gap rising with cut depth and saturating when cuts are combined, and a shallow-compressed layout starts 28 percent lower than compressing every boundary.
Larger model, longer context: where compressed RLVR collapses
A larger base model trained at 16k generation length gained nine points of held-out accuracy by step 150, then fell below its starting level within twenty steps. The trigger was a feedback loop on generation length, amplified by a sampler-trainer gap the compressed path cannot measure. A controlled pair at smaller scale, with its own dense control, shows the same gap is survivable for 600 steps when generations stay short.
Agentic training: where the train-inference gap stops being survivable
The two-circuit method that ran 600 steps without collapse on short single-turn generations was moved to an agentic task, where one trajectory is up to fifty interactions and each is optimised as its own training row. The compressed run stopped completing any task by step six while its dense control held flat. The train-inference gap measures 0.61 against 0.006 dense, and it is already at full size before the first weight update.
Activation compression at inference: PRF and static SVD
A frozen-checkpoint check motivated by the train/inference mismatch. The training circuit keeps 5 percent of boundary activations, so this sweep asks what happens when the same budget is applied during generation across two Qwen3-8B checkpoints.
What compressed generations look like
Recorded continuations for one shared prompt from Qwen3-8B and Qwen3-8B-Base, shown under PRF and SVD at the least aggressive tested rate and at the 5 percent boundary budget.
Communication-efficient GRPO: the system as it stands
The one-page status of the whole system: the two-circuit design, the exact-k PRF boundary codec (77 of 1536 coordinates per token), the sign(M) anchor, dense-level capability in and out of domain after 600 steps at 95 percent compression, and the train-inference mismatch the design lives with.
Same Rank-1, Different Projectability - What Breaks from GSM8K to Big-Math
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.
Rank-1 Trajectories on a Harder Task - Big-Math vs GSM8K
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).
Big-Math signed-EMA (α, β) sweep - how much correction, how fresh an anchor?
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.
Big-Math Training Reward - Does Weight Projection Help?
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.
MOAT Projector Verdict - ANCHOR/FAST from GSM8K to Big-Math
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.
Math-only Projector Ablation
EXP-61 confirms the harder Math trace is only marginally projectable at the freshest anchor and do-nothing-optimal at operational staleness.
Rank-1 Trajectory Family on the EXP-57 Trace
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.
LA-WARMUP-V2 - PP-GRPO at 20/20 Without Collapse
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.