Does late mean near the end of the game?

No. It means later in the move-search order at a particular position.

Is a reduced move completely ignored?

Not necessarily. It is initially searched to a smaller depth, unlike a move that is pruned without that search.

Does deeper re-search eliminate every risk?

No. A move that appears weak in the reduced search may never trigger deeper analysis.

AI Glossary / Graphs and symbolic AI / Intermediate

Late move reduction

Definition

Late move reduction is a selective game-search technique that initially examines some later-ranked moves to a shallower depth than the main search. It relies on move ordering to place promising candidates earlier. By spending less effort on moves expected to be weaker, a search can devote more of its limited computation to other possibilities.

What it means

“Late” means late in the order in which candidate moves are searched at a position, not late in the game or slow to execute. A move may be perfectly legal and potentially strong while appearing late because the ordering heuristic ranked it poorly.
Reduction means shortening the initial lookahead, rather than necessarily skipping the move entirely. This distinguishes late move reduction from late move pruning. The method is a heuristic allocation of search effort: a shallow result can miss a consequence that a deeper search would reveal. It therefore needs to be understood as selective search, not as a universally safe logical elimination rule.

How it works

A game-search program orders legal moves using available information, searches promising candidates, and applies a reduction policy to eligible later moves. The reduction can depend on factors such as search depth, move position, and tactical context. Exact policies vary across engines and versions.
In common designs, a reduced search that indicates a move may improve the current best result can trigger a deeper re-search. That safeguard helps recover promising moves, but it cannot guarantee recovery of a move that looks poor at shallow depth and strong only beyond the shortened horizon. Good ordering, sensible exclusions, and empirical testing are therefore central to the technique.

Three practical examples

A chess engine gives some later-ranked quiet moves less initial depth while investing more search in candidates ranked near the front.
A game-programming experiment compares move quality and node counts with reductions enabled and disabled under the same time budget.
A search debugger records when a reduced move triggers deeper analysis, helping developers understand the effect of their reduction policy.

Think of it this way

Think of giving the most promising proposals a detailed first review while initially screening lower-ranked proposals more briefly. The analogy is limited because a game tree contains adversarial replies and consequences that emerge only after several moves. A brief screen can overlook an excellent proposal, just as shallow search can overlook a strong move.

Worked example

Consider an invented maximizing search node with four candidate moves, A, B, C, and D, in that order. Suppose the normal search would examine each resulting position for five more plies. A ply is one player's move, so five plies is not five complete pairs of moves.
The search of A establishes a current best value, alpha, of 20 in arbitrary evaluation units. A simplified reduction policy searches C and D initially for only three more plies. This is a reduction of two plies from the stated five-ply continuation depth; the example measures depth after the candidate move has been made.
Suppose C returns an estimated value of 10 at the reduced depth. Under the simplified policy, it receives no deeper re-search. D returns 25, indicating that it may improve on 20, so the program re-searches D at the normal continuation depth. That deeper search returns 15, and A remains preferred.
Now suppose C would have returned 40 if examined to the normal depth. The initial reduced result of 10 concealed that opportunity. This possibility explains the risk of the heuristic. The numbers are an abstract search trace, not evaluations from an actual chess position or a description of an exact Stockfish implementation.

Illustrative case study

In a fictional game project, Orchard Engine reduces every move after the first two and reports a large drop in nodes searched. Testing later reveals positions where a quiet defensive move is ranked late and its benefit appears only beyond the reduced horizon.
The developers examine those failures, refine the eligibility policy, and evaluate playing results under fixed time controls. The invented case shows why node savings alone do not establish a stronger search. It does not prescribe particular reduction constants or claim a measured rating improvement.

Common misconceptions and mistakes

Late move reduction is not the same as alpha-beta pruning. Ordinary alpha-beta pruning can preserve the result of a specified full-depth minimax search, while heuristic depth reductions can change which consequences are examined and therefore change the result.
Do not confuse move order with proven move quality. A late-ranked move can be best. Also, a re-search rule is not a guarantee that every overlooked tactic will be recovered, because the shallow score may never trigger that rule. The effectiveness of a policy depends on its interaction with evaluation, move ordering, extensions, and other search mechanisms.

Business use

The direct application is mainly game-engine and adversarial-search engineering. For a team building such a system, define the baseline search behavior and test reductions on representative positions and fixed resource limits. Include difficult cases where a quiet move has a delayed payoff.
Measure more than search volume: compare decision quality, repeatability, and behavior when time is short. Record the exact policy and version used, since small implementation choices can matter. The broader lesson for decision systems is that allocating less analysis to low-ranked options can save effort, but the ranking itself can be wrong and needs evaluation.

Check your understanding

Why can move C still be the best move in the worked example?

Answer: Its reduced-depth value of 10 may hide a stronger continuation visible only at the normal depth. The hypothetical deeper value of 40 would beat A's 20, but the simplified policy does not discover it.

Related terms

Minimax
Ply
Horizon effect
move ordering
alpha-beta search
selective search
late move pruning
principal variation

Learn next

Build a small abstract game tree in which a move's benefit appears beyond a shallow cutoff. Compare full-depth search with a clearly specified reduction policy. Track both saved nodes and changes in the selected move.

Game tree

Sources

Sources: Stockfish, official terminology documentation, Late Move Reductions and Move Ordering: https://official-stockfish.github.io/docs/stockfish-wiki/Terminology.html

Previous: Late Binding Next: Latency