Research map / Reinforcement learning / Visual and representation RL

The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: Scalable RL systems and environments · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Extends the logit lens to Leela Chess Zero's Post-LN policy network, finds correct puzzle solutions that appear in intermediate layers but are overridden at the output ("forgotten puzzles"), and uses concept steering against late-layer safety preferences to recover 61.7% of them.

Problem

Neural networks can internally compute algorithmic solutions (e.g., look-ahead) yet output wrong answers; it is unclear whether failures stem from the algorithm or from competing priors. The paper studies this in a chess policy network with verifiable tactical puzzles.

Approach

A Post-LN logit lens (zero-ablating later sublayer outputs and biases while keeping subsequent normalization) maps intermediate representations of all 64 square tokens to move policies; correct puzzle solutions appear mid-network and are lost late. Replicating look-ahead analyses shows the algorithm works on these positions, and gradient steering against late-layer safety concept preferences recov

Key results

• Gradient steering against opponent threats at layer 13 (alpha = -0.5) raises solve rate from 88.6% to 92.4% on the general puzzle set and recovers 61.7% of forgotten puzzles, with 127 regressions.
• Of 7,901 forgotten puzzles, 88.6% involve sacrifices versus 26.4% of solved puzzles, and the model's value head favors the correct-move position in 95.8% of cases.
• Single attention entry in L12H12 ablated reduces correct-move log-odds by more than 1.5 in over 10% of forgotten puzzles; bilinear probes decode the third move at 80% accuracy at layer 12.

Surprising findings

• Correct solutions, including mate-in-one and mate-in-two, are discovered at intermediate layers yet overridden at the output, even though look-ahead representations remain causally active and linearly decodable.
• Probes on randomly initialized models also show lower accuracy on forgotten puzzles, complicating the probe-based evidence.

Limitations

• Causal analyses are conducted on a single Leela network; the forgotten puzzle phenomenon is only verified across other networks.
• Safety priors are identified via Stockfish evaluation terms as proxies, not a precise mechanistic account of how priors are implemented.
• Steering yields regressions (127 puzzles) and gains are measured with a first-order gradient approximation.

Experimental setup

Datasets: Lichess puzzles (general dataset, 10,000 puzzles) CCRL positions Encyclopedia of Chess Openings positions

Benchmarks: Lichess puzzles general dataset

Metrics: solve rate Elo rating log-odds of correct move Kendall tau policy entropy concept preference (centipawns)

Baselines: unsteered full model random steering directions Leela policy network anchor

Builds on: logit lens Tuned Lens DeepNorm AlphaZero learned look-ahead in Leela Chess Zero

Closest papers by meaning

• Evidence of Learned Look-Ahead in a Chess-Playing Neural Network (NeurIPS 2024 · similarity 0.92)
• Implicit Search via Discrete Diffusion: A Study on Chess (ICLR 2025 · similarity 0.88)
• Amortized Planning with Large-Scale Transformers: A Case Study on Chess (NeurIPS 2024 · similarity 0.88)
• Human-Aligned Chess With a Bit of Search (ICLR 2025 · similarity 0.87)
• PUZZLES: A Benchmark for Neural Algorithmic Reasoning (NeurIPS 2024 · similarity 0.87)
• Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners (NeurIPS 2026 · similarity 0.86)