- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
In this paper, authors propose an execution contract for removing physical model rows in verifier-guided sequential Monte Carlo while preserving the logical particle state that determines selection. It separates logical slots from physical rows, retains parked weights for terminated particles, defines unit-barrier and eager-prune schedules, and gives a consistency and finite-N equivalence result under append-stable segmentation, absorbing zeros, deterministic verification, a shared barrier clock, slot-indexed randomness, and batch-independent proposal kernels.
Critical strengths
The logical state separation and the requirement to retain the selection weight of an evicted but still eligible particle address a subtle failure mode directly and concretely. The execution contract is unusually explicit about exactly-once score application, barrier-clock behavior, compaction, and random-stream handling. The work savings from finished-row eviction are large and consistently observed across several benchmark families.
Critical weaknesses
The paper measures substantial batch-shape-dependent logit changes under default kernels, while the fully invariant reference uses per-row normalization/attention and a sequential sampling scan whose overhead at the studied particle widths is not measured. Consequently, the strongest finite-N identity result is demonstrated for an audit-oriented reference rather than a practical high-throughput implementation, and default-kernel results are only coupling diagnostics rather than a distributional guarantee. The empirical scope is also narrow (one 1.5B model, mostly N=8, a custom Hugging Face runner, and no capacity-matched production runtime); throughput comparisons have unequal peak occupancy, and hidden-quality effects are underpowered and unresolved. Finally, timeout-containing HumanEval runs do not meet the paper's own pure-verifier premise, so the formal guarantee does not cover every reported workload cell.
Questions for authors
Refer to Weaknesses.