Hand-set neural networks master small games but struggle to turn tactics into plans

Programming weights instead of learning them

The November 2026 preprint studies “weight-space programming”: explicitly assigning a neural network’s numerical weights to implement a specified computation, rather than adjusting them through gradient descent on training examples. The resulting game players select moves through a fixed forward calculation, without running a search during play.

The construction process was considerably less restricted. The main experiment used a coding agent with code execution, iterative testing and exhaustive-search tools. Previously solved game positions could be calculated offline and embedded in the network as constants. Thus, “zero-shot” here does not mean that every network emerged from one unaided attempt. A separate cold, tool-free evaluation protocol is proposed for future use.

The networks combine simple mathematical components that act like logical gates, recognize exact board positions and assign priorities to candidate moves. Immediate wins, defensive moves and positional preferences can receive different scores. Where those rules fail, stored exceptions can override them.

A ladder of increasingly demanding games

The study reports ten networks across paired tasks: Nim and Wythoff’s game; tic-tac-toe and Dots and Boxes; reduced and standard versions of Connect Four and Othello; and chess endgames with a king and rook, or king and queen, against a lone king.

The pairs are intended to distinguish recurring difficulties from quirks of individual games. The proposed difficulty dimensions include state-space size, availability of a compact strategy, move geometry and the length of planning required.

For six bounded tasks, the author reports exhaustive checks against separately written exact solvers, rather than estimates from sampled games. These include the reduced Connect Four task, actually a four-column, five-row connect-three game, with 104,399 nonterminal positions, and reduced Othello with 53,651 nonterminal positions under the study’s counting convention.

These successes do not all reflect equally compact reasoning. Nim uses an arithmetic strategy. Other networks rely partly or extensively on stored solutions. Exact play therefore demonstrates successful encoding, but does not necessarily demonstrate discovery of a short general strategy.

Strong tactics, weak endgame plans

The chess experiments make the distinction between local correctness and sustained planning particularly clear. Both main networks always selected legal moves across their exhaustively checked legal positions. Yet that guarantee did not translate into reliably finishing games.

The king-and-rook network took an available immediate checkmate 84.1% of the time. Across 3,000 simulated games against a worst-case adversarial defender, it achieved checkmate in only 0.57%. Most games instead ended with the rook lost or the move limit exhausted.

The king-and-queen network included an exact detector for immediate checkmate and stronger safety checks. It found every available immediate mate in the exhaustive audit and avoided avoidable stalemates and unforced queen captures. Nevertheless, it checkmated in only 9.13% of 3,000 adversarial games. Most remaining games reached the move cap.

The queen construction changed both the piece and the safety machinery, so its improvement cannot isolate planning length as the sole cause. It does show that reliable immediate decisions can coexist with poor completion of a longer objective.

Search opponents expose another gap

On standard boards, the main Connect Four network won 99.5–99.8% against random play, depending on seating. The Othello network won 86.3–87.5%. Each recorded no wins in 800 games against its respective independently written search opponent, looking four plies ahead in Connect Four and three in Othello. A ply is one player’s move.

Those are substantial performance gaps, but not a clean experiment changing search alone. Although parts of the paper describe opponents sharing the networks’ evaluation weights, its detailed accounts say the opponents used independent evaluations. That inconsistency weakens the claim that the comparison isolates an architectural cause.

What the replications clarify

Additional construction attempts broadly reproduced exact solutions on small tasks and difficulty with harder planning. They also exposed important qualifications: differing state-count conventions, a disputed Dots and Boxes input ambiguity, unequal opponents and sample sizes, and some chess implementations that did not satisfy the weight-only requirement. Some replicated networks retained substantial win rates against search, contradicting any blanket claim that shallow search always defeats these heuristics.

Most importantly, one replication embedded exact chess endgame solutions computed offline. Those networks forced mate throughout the checked legal positions. The architecture can therefore represent successful long-term policies when supplied with a solved table.

The defensible conclusion is narrower than a fundamental barrier to neural-network planning: these constructions often encoded rules and tactics successfully, but their compact hand-derived heuristics struggled to supply a complete strategy. The preprint does not establish that all feed-forward networks, or all language models, share that limitation.