Draft
Produce a complete first attempt.
A pinned language-model owner receives the problem and writes one coherent internal trajectory.
-50+5+10+15+20+25+30+35
Inspect
00.00100.00ORIGINLIVE FIELD100M300M1.0B3.0B10B30B100B200B
Independent language-model research
Project Shohin
A model should be able to revise its own thinking.
Shohin separates generation into learned temporal roles: draft a solution, revise that internal trajectory, then commit one coherent answer. The goal is better reasoning from the same underlying model—not another larger model.
Current experiment
Temporal revision improves dense hosts from 0.8B through 9B and now has measured MoE gains on Qwen3.6-35B-A3B and Mixtral-8x22B. Mixtral gains 301 answers over unchanged and 92 over matched self-refinement, while its capability-preservation gate remains unresolved.
See the scale boundary →Capability field
Measure the lift.
Map the frontier.
Switch between the historical matched-board field and the current official Qwen3.5-9B benchmark campaign. Five verified ledgers are plotted now; unfinished boards remain visibly pending.
Benchmark viewAverage improvement
01
Current architecture
Reasoning unfolds across learned roles.
Standard decoding commits each token while the solution is still being formed. Shohin preserves a complete first trajectory and turns revision into an explicit, trainable computation stage.
Revise
Train a later owner to edit it.
A same-family revision owner sees the source and draft, then emits a complete replacement—not a patchwork of hidden states.
Commit
Select one intact trajectory.
A learned policy compares the two answers and commits exactly one. No task label, correctness bit, tool, or verifier is available at inference.
problem→draft→revision→one answer
02
Measured capability
The extra pass is not the result. Training it is.
An identical untrained second pass solves 316 protected problems. Training the revision owner raises that to 374; whole-trajectory commitment reaches 383 while preserving executable-code accuracy.
- Matched second pass
- 316 of 538
- Trained revision
- 374 +58 answers
- Final commitment
- 383 +67 answers
- Macro movement
- +8.552 percentage points
| Model | Scale | Checkpoint | Macro | GSM8K | MATH-500 | Executable code | GPQA | BBH logic | AIME 2024 |
|---|---|---|---|---|---|---|---|---|---|
| Shohin · draft/revise/commitConfirmed product result ↗ | Qwen3.5-9B · same-family owners | 383 / 538 solved · protected board | 75.815% | 87 / 100 | 72 / 100 | 35 / 40 | 114 / 198 | 75 / 100 | 6 / 30 |
| Trained internal revisionRevision result ↗ | Qwen3.5-9B · IDR1 | 374 / 538 solved | 75.005% | 88 / 100 | 69 / 100 | 35 / 40 | 104 / 198 | 78 / 100 | 3 / 30 |
| Matched original second passMatched control ↗ | Qwen3.5-9B · B1 | 316 / 538 solved | 67.263% | 85 / 100 | 60 / 100 | 34 / 40 | 62 / 198 | 75 / 100 | 4 / 30 |
| Coherent trajectory oracleSelection ceiling ↗ | Not deployable · upper bound | 399 / 538 solved | 78.619% | 90 / 100 | 77 / 100 | 35 / 40 | 118 / 198 | 79 / 100 | 6 / 30 |
The result uses a Qwen3.5-9B host and does not establish frontier parity, unrestricted reasoning, or the same capability in the original 125M scratch model. The coherent oracle is diagnostic, not deployable.
Read methods, controls, and failures03
Research standard
Claims should survive the obvious alternative explanation.
Shohin keeps mechanism tests, product benchmarks, and historical failures separate. The public record includes the reason an experiment was rejected—not only its best number.
Matched controls
The comparison receives the same model family, source, internal draft, second pass, decoding budget, and evaluator.
Source-disjoint evaluation
Training and evaluation identities are separated before fitting. Protected results are opened only after the gate is frozen.
Failures stay public
A higher average is rejected when a protected domain regresses. Closed lanes remain in the research archive with their controls.
Does the architecture improve a capable language model?
Measured on broad natural-task boards with matched model controls.
Evidence →Which internal state actually causes the answer?
Tested in controlled worlds using source deletion and destructive interventions.
Architecture →04
Open record
The path matters as much as the result.
Preregistrations, controls, failed gates, corrections, and source artifacts remain available in the complete research archive.