Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.25963

ABACUS: Program-Guided Multistep Reasoning over Frozen Judgment Models

Yuhao Wang·Zelin Zang
View original source ↗Read source document ↗Export metadata ↓

Abstract

Judgment models answer semantic questions by scoring a finite set of candidate outcomes. Can reusable programs turn these local decisions into powerful multistep reasoning? We introduce \method, a framework that jointly constructs judgment schemas and composition programs for frozen judgment models. A generative teacher designs the procedure offline; the judgment model resolves local semantic questions online, and the saved program composes its decisions into an answer. We study three allocations of design effort between humans and the teacher, focusing on construction through a shared interface. We instantiate the framework with Kev. On three selected BIG-Bench Hard tasks, nine independently constructed programs solve all 450 executions of 150 development inputs. A separate deployment study freezes three programs after development-set reference-agreement refinement and evaluates them on 600 held-out questions. These programs achieve near-perfect 99.67\% macro accuracy, versus 64.00\% for direct Kev and 68.80\% for final-answer-only Qwen3.5-4B. They attain comparable observed accuracy to Qwen's unrestricted response policy (99.07\%) with an \textbf{8.43-fold online speedup}: 0.255 versus 2.150 seconds per question. These results show how reusable programs can substantially expand the reasoning capability of a compact judgment model while keeping its parameters fixed.

Explore related subjects

cs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yuhao Wang, Zelin Zang. 2026-10-19. ABACUS: Program-Guided Multistep Reasoning over Frozen Judgment Models. https://arxiv.org/abs/2610.25963

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Agents-as-Peers: Bias-Controlled LLM Evaluation as Multi-Agent Peer Assessment

Large language model (LLM) evaluation commonly relies on fixed human-authored benchmarks or centralized human/model judges. Fixed benchmarks can age and enter model training corpora, while centralized judges introduce their own scoring preferences. We cast LLM evaluation as a multi-agent peer-assessment problem and introduce Agents-as-Peers, a distributed evaluation protocol in which every agent generates tasks, answers the shared task set, and judges the answers of its peers, with no privileged evaluator. Self-ratings are excluded from the ranking, and judging is repeated under controlled identity and presentation regimes so that self bias, identity effects, presentation-position sensitivity, and judge severity are measured rather than left implicit. In an instantiation with 12 LLMs and 420 agent-generated questions, Agents-as-Peers aggregates 55,440 non-self peer judgments per regime. Most agents assign their own answers higher scores than those answers receive from peer judges; visible identity changes scores for some agents; the first fixed answer slot is associated with scores 0.39 points higher than the shuffled baseline; and judges differ markedly in severity. On labeled benchmarks, peer scores align strongly with accuracy on TruthfulQA (Pearson r=0.904) and GSM8K (r=0.873). On TruthfulQA, the peer-score correlation is substantially higher than the self-score correlation (r=0.538). A benchmark-overlap screen and item-level human audit further characterize the endogenous task set. These results support Agents-as-Peers as a cohort-relative measurement protocol whose ranking is reported together with the biases and rater effects that shape it. Code and dataset: https://github.com/caura-ai/caura-PeerRank

cs.AI↗
arXiv2026

From Debate to Decision: Conformal Social Choice for Safe Multi-Agent Deliberation

Multi-agent debate improves LLM reasoning, yet agreement is not evidence of correctness. When agents converge on a wrong answer through social reinforcement, consensus stopping commits that error to automated action with no recourse. We introduce Conformal Social Choice, a post-hoc layer that converts debate outputs into calibrated act-versus-escalate decisions. Verbalized probability distributions from heterogeneous agents are aggregated via a linear opinion pool and calibrated with split conformal prediction, yielding fixed-round marginal coverage: at any pre-specified round, the correct answer is included with probability ${\geq}\,1{-}α$, without assumptions on individual model calibration. A hierarchical action policy maps singletons to autonomous action and larger sets to human escalation. Across eight MMLU-Pro domains and two additional benchmarks, fixed-round coverage stays close to target. The key finding is not that debate becomes more accurate, but that the conformal layer makes its failures actionable: at the final round, 81.9\% of wrong-consensus cases are flagged by non-singleton sets at $α{=}0.05$. Separately, among cases that reach a singleton before the final round, accuracy reaches 90.0--96.8\% on MMLU-Pro (up to 22.1pp above consensus stopping): a selection effect, not a reasoning improvement. Simple probability- and entropy-threshold rules attain comparable interception; conformal's distinct value is a finite-sample fixed-round coverage guarantee together with explicit candidate sets for escalation. Adaptive first-singleton stopping is evaluated empirically and not covered by this guarantee.

cs.AI↗
arXiv2026

Non-Stationarity Breaks Permutation Surrogates in Multi-Agent Reinforcement Learning: Diagnosis and Remedies

Reporting guidance for information-theoretic measures is rarely tested against ground truth. We test one guardrail in two multi-agent reinforcement learning games, a social dilemma and a coordination race, where directed influence between selected agent pairs is zero by construction, over 100 seeds. Omitting one precondition, exclusion of the non-stationary training transient, gives false-positive rates of 100.00% and 99.95%: agents annealing exploration independently, in runs that never met, are flagged as influencing one another. Excluding the transient reaches 3.05% in the social dilemma but 11.8% in the coordination game, so the non-stationarity must be treated and exclusion is neither the only way nor sufficient. What we recommend instead changes the null model. Series, statistic and estimand are left untouched: permuting the source within blocks of training time reaches 5.25% and 5.50%, the only one of the three at or below the size of the test in both games. The construction is not new, being a restricted permutation local to blocks of training time; the measurement is. A third null control, four learners in the social dilemma, defeats all three, and there conditioning costs three times what ours does, 15.20 points against 4.55. Titrating injected links separates it from conditioning on identical pairs, 82.5% against 58.0%, while the ablated test reports a positive with or without a link, so its apparent sensitivity is uninformative. Five block counts spanning 16 to 256 land in the nominal region. Code and data are released.

cs.AI↗