Paper Citation Record · LEDGER

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling

As of 7 August 2026, this Paper Citation Record lists 13 of 13 outbound references and 2 inbound Pith citation observations for arXiv:2606.12370.

A citation records a reference. It does not transfer a finding from one paper to another.

pith.paper-citation-record.v1
2606.12370 v1

Coverage vector

measured 13 of 13 reference resolution

Typed states for the displayed outbound observations.

Source: paper_references, paper_reference_links, observed 2026-07-11T11:50:26.030339Z

measured 15 of 15 standing notices

One-hop event checks from named stored sources.

Source: scholarly_work_events, retraction_status_cache, observed 2026-08-07T06:34:17.273281+00:00

measured 2 of 2 inbound itemization

Pith citing papers itemized under the disclosed page cap.

Source: paper_references, paper_reference_links, observed 2026-08-02T01:34:48.709434Z

measured 0 of 1 external citation measurements

A source-named dated measurement, never combined with another source.

Source: cited_works

Reference resolution

13 of 13 outbound references displayed

verified exact1
verified fuzzy0
unresolved12
parse uncertain0
malformed identifier0
metadata mismatch0

External citation measurements

No source-named external measurement is stored.

Outbound references

Observation 7a1fdafe-eeae-436e-867c-e1b7bb751525 · outbound

This paper cites Draft-OPD: On-Policy Distillation for Speculative Draft Models.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling Draft-OPD: On-Policy Distillation for Speculative Draft Models

Reference 1

Resolution: verified exact; local_arxiv, observed 2026-06-27T10:30:51.814090Z

Source-reported events for the cited work

No event found in the named queried sources as of 2026-08-07T06:34:17.273281+00:00.

Observation 5d0b5a34-4160-4c2e-ad9a-8c1252c72cc3 · outbound

This paper cites The TV gradient is proportional to qj, so it automatically ignores low-probability tokens.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling The TV gradient is proportional to qj, so it automatically ignores low-probability tokens

Reference 2

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 358997f3-c559-44c6-a2f9-0f2a1e876189 · outbound

This paper cites rejected under rejection sampling.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling rejected under rejection sampling

Reference 3

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation a971f4a7-890b-4a1b-887a-74b13de80b7c · outbound

This paper cites The TV gradient is bounded byq j.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling The TV gradient is bounded byq j

Reference 4

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 8158ee0a-c955-4abd-95b2-88fff7cb35bf · outbound

This paper cites mode-seeking.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling mode-seeking

Reference 5

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation de69c6a3-4937-41a7-8aac-f7c89221e1d3 · outbound

This paper cites The reverse KL imposes anasymmetric 22 penalty: over-estimation (qj >p j, so log(qj/p j)> 0) incurs a much stronger gradient than under- estimation.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling The reverse KL imposes anasymmetric 22 penalty: over-estimation (qj >p j, so log(qj/p j)> 0) incurs a much stronger gradient than under- estimation

Reference 6

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation d381fef1-2418-4f18-baab-034185babcf3 · outbound

This paper cites an unresolved cited work.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling Unresolved cited work

Reference 7

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 3f4f813b-9737-403a-8f47-1ee4b3bef279 · outbound

This paper cites an unresolved cited work.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling Unresolved cited work

Reference 8

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation ac4e1006-56f6-4b4b-8a67-1cb43cd31550 · outbound

This paper cites Practical considerations.The above analysis assumes δ is a constant, but in practice, the draft head has finite capacity.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling Practical considerations.The above analysis assumes δ is a constant, but in practice, the draft head has finite capacity

Reference 9

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation bf0256dc-d046-46a2-85d2-54be6555b236 · outbound

This paper cites , γ, draw ui ∼Uniform( 0, 1) and accept ˆyi if ui ·q i( ˆyi)<p i( ˆyi), i.e., with probability min 1,p i( ˆyi)/qi( ˆyi).

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling , γ, draw ui ∼Uniform( 0, 1) and accept ˆyi if ui ·q i( ˆyi)<p i( ˆyi), i.e., with probability min 1,p i( ˆyi)/qi( ˆyi)

Reference 10

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 8d99502d-fb3f-4b77-afa6-dccfd14566c4 · outbound

This paper cites For rejection at step j, the residual distribution is presid(v)∝max( 0, pj(v)−q j(v)); for the bonus token (all accepted), the residual is simply pγ(v).

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling For rejection at step j, the residual distribution is presid(v)∝max( 0, pj(v)−q j(v)); for the bonus token (all accepted), the residual is simply pγ(v)

Reference 11

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 18c4527f-6d18-42e6-a7cf-84e12de6c290 · outbound

This paper cites The kernel records the index of the first rejected step.

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling The kernel records the index of the first rejected step

Reference 12

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation 07634fa9-0f6b-4b60-945c-8a00a5d2d8c4 · outbound

This paper cites For rejection at step j: zresid(v) =log max( 0, pj(v)−q j(v)); for the bonus token: zresid(v) =z target,γ(v) (the raw target logits).

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling For rejection at step j: zresid(v) =log max( 0, pj(v)−q j(v)); for the bonus token: zresid(v) =z target,γ(v) (the raw target logits)

Reference 13

Resolution: unresolved; no resolver link, observed 2026-06-27T10:24:34.257293Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Pith citing papers

Observation 7b39c9c3-e86a-4261-9322-81a945c9f801 · inbound

D-cut: Adaptive Verification Depth Pruning for Batched Speculative Decoding cites this paper.

D-cut: Adaptive Verification Depth Pruning for Batched Speculative Decoding Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling

Reference 82

Resolution: unresolved; no resolver link, observed 2026-08-02T01:34:48.709434Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.

Observation b1262a26-beb5-4bf9-b512-c4321cde853f · inbound

AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding cites this paper.

AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling

Reference 17

Resolution: unresolved; no resolver link, observed 2026-08-01T01:22:13.804729Z

Source-reported events for the cited work

Unavailable: canonical work link unavailable.