Looped refiners swap their per-pass target: multi-answer Sudoku jumps from 2.5% to 98.8%, with no answer list needed
Synopsis
Looped refiners apply one small network repeatedly to a draft answer, but deep supervision trains every pass toward the single answer the dataset stored; when a puzzle has many correct answers, that target blends them into an invalid grid and leaves a loss floor no model capacity removes. Training each pass instead toward the correct answer nearest its own input, a proximal target, lets the same 1.73M-parameter network solve 98.8% of Sudoku puzzles with 2–64 answers, against 2.5% for the standard target, and 93.8% of sparse puzzles with more than 10,000 answers and no answer list, against 0.0%.
Interpretation
The per-pass training target of a looped refiner moves from the one answer the dataset stored to the correct answer nearest that pass's own input, and solving on multi-answer puzzles rises from single digits to above ninety percent. Looped refiners such as HRM and TRM previously used deep supervision to train every pass toward one stored answer per puzzle; on Sudoku with 2–64 answers the standard target solved 2.5%. With the same 1.73M-parameter network and 5 seeds, the proximal target solves 98.8% of Sudoku puzzles with 2–64 answers against 2.5% for the standard target; making the network 4 times larger, at 6.9M parameters, leaves the standard target at 2.4%.
A pass only has to make fewer errors than its input had for the loop to converge, and that pass's training loss bounds its error count. Under the standard target the loss-minimising pass ignores the draft and splits probability across answers in each cell, decoding to a blended grid rather than an answer. On 1,000 multi-answer test puzzles with 4 starts each and 3 seeds, the proximal model leaves 0.28–0.49 times as many errors as its input in passes 1–6, while the standard target exceeds 1 from pass 3 and stalls 6.7 cells from the nearest answer.
The target can come from a complete answer list, from a depth-first solver started at the draft, or from a quality band set by a knob, and the latter two need no pre-listed answers. When only some answers are listed, the nearest listed answer depends on which answers were kept, and blending partly returns. With no answer list and targets from a solver started at the draft, sparse Sudoku solves 93.8% (standard 0.0%) and 200-node 3-colouring solves 90.9% (standard 36.9%).
Perspective
The demonstrated setting is Sudoku, sparse Sudoku, planted 3-colouring, the multi-solution benchmark of Nandwani et al. and a Scrabble board, with models of at most 6.9M parameters and training that needs a complete answer list, a solver that can start from a draft, or a quality ranking.
The loop finds 53.6% of a Sudoku puzzle's answers from the same starts, below its target's 64.3%; the Scrabble knob is useful over 0.5–1.0, and the Sudoku knob was learned in 2 of 3 seeds; on N-Queens the model solves 99.9% of puzzles from seen solution families but 2.7% from held-out ones.
