Skip to main content

# Policy replay

2 related articles

Subscribe to updates

arXiv Oct 8, 2026
MARS compares direct classification with claim scoring on 1,195 PE/ELF binaries: direct classification is more accurate across all six models, while retained claims enable policy replay
MARS is a malware triage framework for PE and ELF binaries that, with a shared evidence collector and identical static evidence bundles, compares direct LLM classification against structured behavioral claims scored by a fixed policy; across 1,195 binaries, 1,001 near-duplicate clusters, and six models, direct classification is more accurate for all six models with paired advantages of 3.7 to 20.9 percentage points, while the claim path retains the policy inputs and permits replay and policy revision without another model call.
MARS is a malware triage framework for PE and ELF binaries that, with a shared evidence collector and identical static evidence bundles, compares direct LLM classification against structured behavioral claims scored by a fixed policy; across 1,195 binaries, 1,001 near-duplicate clusters, and six models, direct classification is more accurate for all six models with paired advantages of 3.7 to 20.9 percentage points, while the claim path retains the policy inputs and permits replay and policy revision without another model call.
Hyeongjun Choi（First author, Corresponding author）
MARS Malware Triage Structured claims Verdict policy Policy replay
arXiv Sep 29, 2026
Freezing debate prevented 29 collapses but lost 108 corrections across 6,925 MMLU-Pro debates
The work introduces an auditable protocol for homogeneous three-agent LLM debate that decomposes final accuracy into preserved, collapse, correction, and unrepaired transitions with collapse onset and signed intervention utility, identifies 253 collapses across 6,925 MMLU-Pro debates, and shows via replay that a leave-one-model-out probe-gated freeze prevents 29 collapses while losing 108 corrections under equal weights, so collapse prevention alone can recommend the wrong policy.
The work introduces an auditable protocol for homogeneous three-agent LLM debate that decomposes final accuracy into preserved, collapse, correction, and unrepaired transitions with collapse onset and signed intervention utility, identifies 253 collapses across 6,925 MMLU-Pro debates, and shows via replay that a leave-one-model-out probe-gated freeze prevents 29 collapses while losing 108 corrections under equal weights, so collapse prevention alone can recommend the wrong policy.
Nanyang Technological University
Xin Li（First author）
Multi-agent LLM debate Transition ledger MMLU-Pro Policy replay

You’re all caught up