Abstract
We introduce POND (Prefill-Only, No-Decode), an architecture that separates large-model prefill from small-model decoding. A frontier-scale sender—Qwen3.8 Max(2.4T), GLM 5.3(753B), or DeepSeek-V4-Flash(284B)—processes the input without generating text. A learned bridge transfers its hidden states as additional attention keys and values to Qwen3.5-4B, a compact receiver capable of running on suitable edge hardware. The receiver generates the complete answer using both the original prompt and the latent guidance. Optional periodic review lets the sender read the partial response and append updated guidance while preserving the receiver’s existing KV cache. Both base models remain frozen; only the bridge is trained.
Selected successful reviews
Inference Examples
A single review can catch a local error, replace a flawed approach, or bring a runaway derivation back to a verified answer.
57/824B baseline
→
69/82with one GLM draft
→
70/82best reviewed setting
01Local correctionAIME 2022 · 026
One off-by-one check changes 73 to 72
Connect two of the points numbered 1–20 when their difference is prime. How many triangles are formed?
Before review73For the twin-prime case p = 11, the span was mistakenly treated as 12, allowing eight starting points.
After one review72The review catches that the span is p + 2 = 13, so a ≤ 7. Both cases contribute 36.
Improvement Pinpointed and repaired a single off-by-one error · incorrect → correct
Show full responses
02RecoveryAIME 2023 · 037
A runaway search converges to a verified answer
Find the largest prime p < 1000 satisfying the Gaussian-integer and triangle conditions in the problem.
Before reviewNo answerThe derivation loops while testing whether 977 is a sum of two squares, producing 221,277 characters without a reliable conclusion.
After one review349With z = 18 + 5i, it obtains x = 4482, y = 4735, and verifies |x − y| = 253 < 349.
Improvement Restored a stalled derivation and reduced output from 221k to 24k characters · incorrect → correct
Show full responses
Before review · 221k characters
Loading…
After review · 24k characters
Loading…
03Method revisionAIME 2023 · 047
An incorrect DP gives way to a short counting argument
Place 1–12 in a 2 × 6 grid so adjacent numbers never differ by a multiple of 3. If there are N arrangements, find the number of divisors of N.
Before successful review56 / 546Complicated state classifications and a faulty dynamic program produce inconsistent, incorrect counts.
After review144Each residue is omitted by exactly two columns: 6!/(2!2!2!) = 90. Two orientations give 180 residue grids, then N = 180(4!)³.
Improvement Replaced a brittle DP with a compact combinatorial invariant · incorrect → correct
Show full responses
04Historical review exampleOpenR1-Math-220k
Correct factor structure replaces a broken enumeration
What is the largest k for which 311 is a sum of k consecutive positive integers?
Before review2048The factor list contains arithmetic errors and even treats non-divisors as factors of 2 · 3¹¹.
After review486From k(2n + k − 1) = 2 · 3¹¹, the largest valid smaller factor is 2 · 3⁵ = 486; the first term is 122.
Improvement Rebuilt the factor enumeration and produced a directly verifiable construction · labeled improved
Show full responses
These are selected qualitative successes. The aggregate 82-task results above include all evaluated examples across AIME, MBPP, HumanEval, and LiveCodeBench.