Optimizer Card — mcts-lats v1.0.0
Szepesvári 2006) extended with LATS (Language Agent Tree Search, Zhou et al.
9 sections3 min read
On this page
1. Identity#
Optimizer name: mcts-latsAlgorithm family: Monte Carlo Tree Search (UCT, Coulom 2006; Kocsis &
Szepesvári 2006) extended with LATS (Language Agent Tree Search, Zhou et
al. 2023) — value-augmented selection.Implementation source: libs/nous/agreement-search/src/mcts-lats.ts with
shared GA primitives in ga-operators.ts.Intended use: trajectory-interpretable search where the workbench must
replay the path from root to winning leaf as an explicit clause- edit sequence
("this is how we arrived at the agreement"). Also the natural integration
point for §179.9.4.2 mediator-value models.
2. Scoring rule#
Utility aggregation: scalar fitness Q(c) per node, caller-supplied via
FitnessFn.Selection score:
score(c) = (1 − λ) · Q(c) + λ · V(c) + C · sqrt(ln N(p) / N(c)).
λ = valueWeight. λ = 0 recovers plain UCT; λ = 1 recovers pure
value-function rollout. C is the UCB exploration constant.Fairness metrics reported: caller selects; full fairness-metrics.ts set
attaches to the run summary.Handling of hard constraints: every node passes through
filterCandidates; failing nodes are not added to the tree.Handling of uncertainty: rollout fitness can be posterior mean, LCB, or
Thompson; the kernel is agnostic.
Candidate space: sandboxed clause-mutator DSL + recombination borrows from
a reference pool.Initial population / seeds: root from §179.4.1.1; reference pool from
§179.4.1.2 decomposition.Time / iteration budget: iteration count + wall-clock cap + rollout depth.Randomness: seeded mulberry32.
4. Outputs#
Accepted candidate contract: highest-Q leaf along the recorded path from
root.Pareto-frontier diagnostics: not the goal; route to NSGA-II / MAP-Elites
for frontier coverage.Trajectory record: the path from root to leaf as an ordered list of
(action, child-id, Δ-fitness) tuples — consumed by the workbench audit
replay.Uncertainty propagation: leaf carries posterior detail; tree visit-count
distribution is reported for confidence visualization.
5. Evaluation#
Metric
Value
Evaluator card
Date
Trajectory faithfulness
replay byte-identical for given seed
n/a — mcts-lats.test.ts reproducibility assertion
2026-04-25
Diversity preservation
not the goal here
n/a
2026-04-25
Fairness (Nash / KS / max-min)
reported per leaf; gate runs against
fairness-suite 2026-04-25
§179.10.6 baseline
Regret vs oracle
bounded by UCT regret
baseline-benchmark-gate 2026-04-25
(O(sqrt(ln N · N)))
Specification-gaming resistance
spec-gaming-audit.ts on winning leaffairness-suite 2026-04-25
Runtime (typical)
sub-second to seconds for 50 clauses;
n/a — deterministic synchronous kernel
2026-04-25
interruptible at any iteration
Cost per run
$0 LLM-free; with valueEstimate
n/a — orchestrator scoring-cost trace
2026-04-25
callback the cost is value-model cost
6. Known limitations#
Tree size grows with iterations; memory is O(nodes × clauses). Practical
cap: ~100k nodes for 50-clause cases.
Without a value model (valueWeight = 0), MCTS falls back to UCT; convergence
on adversarial landscapes can be slow.
Rollout depth and exploration constant C are caller-supplied;
misconfiguration yields shallow or biased exploration.
7. Guardrails#
Candidate safety gate: filterCandidates on every node.Clause static validation: inside filterCandidates.Redline separation: hard constraints filtered before tree insertion.Trajectory audit: trajectory record is part of the §179.5.3.3
reproducibility manifest.
8. References#
libs/nous/agreement-search/src/mcts-lats.tsCoulom (2006), "Efficient selectivity and backup operators in Monte-Carlo tree
search".
Kocsis, Szepesvári (2006), "Bandit based Monte-Carlo planning".
Zhou et al. (2023), "Language Agent Tree Search Unifies Reasoning, Acting, and
Planning in Language Models" (LATS).
docs/research/bibliography.md §tree-search-reasoning