Searcharxiv⌕ Search

arXiv subjects

Junru Zhu

Publications and source records attributed to Junru Zhu.

8 recordsLinked to original sources

SwarmSafe-Delta: Isolating Message Effects on LLM Agent Safety

When a tool-using LLM agent team violates policy, aggregate outcomes do not reveal whether an upstream message changed the executor's action. We introduce SwarmSafe-Delta, a paired executable protocol that freezes upstream computation, varies real, hidden, and format-matched placebo delivery, and scores a single committed action against deterministic benign and risky policy states under preassigned budgets. Across four models and constructed optional tasks, real delivery at $N=4$ produces 21.4% risky-world violations, versus 5.0% with placebo and 1.0% with hidden delivery (paired differences +16.4 and +20.3 percentage points). Real delivery also improves benign task success, illustrating that suppressing communication sacrifices utility. In a second regime constructed to require an upstream lookup route, real delivery improves safe completion by 74.2 points relative to hidden delivery. These effects quantify message-delivery contrasts under specified information dependencies, not naturally occurring failure rates. Exploratory action-level analyses show that the effect depends strongly on whether the upstream message recommends a prerequisite check or direct execution; model differences also largely track these message compositions. The contribution is a controlled, execution-grounded measurement protocol rather than a general claim that agent communication is harmful or protective.

cs.MA↗

Stress-Testing Structure-Aware Calibration of Malware Graph Neural Networks under Type Shift

Post-hoc malware calibrators can condition confidence on graph structure, but their structural inputs may leave the support represented by validation data under malware-type shift. We study this risk on MalNet-Tiny by holding out each of four malware types across three seeds, freezing a graph isomorphism network, and fitting post-hoc mappings only on known-type data. On known-type tests, generic and combined structural calibrators have negative log likelihood (NLL) of 0.4341 and 0.4628. Their mean held-out-type gap widens to 0.1369, with a four-type 95% interval of [-0.0125, 0.2864] and improvement in one of four types. Eight label-independent nuisance variables produce a smaller mean NLL increase of 0.0415, but the paired combined-versus-nuisance interval includes zero. We then use calibration data to define a community-support guard: predictions outside the 95th percentile of calibration-standardized community displacement revert to the generic calibrator. The guard reduces combined NLL by 0.1133 and high-confidence errors from 43.25 to 33.25 per 400-sample cell. It also outperforms matched-rate random fallback on the original folds, while generic and uncalibrated probabilities retain lower held-out-type NLL. This exploratory recovery localizes harmful combined predictions; the stress test exposes heterogeneous calibration failures without establishing a general structural-calibration penalty or a deployment gain over simpler alternatives.

cs.CR↗

Token-Budgeted Escalation for Financial Document Question Answering: Cost Is Predictable, Benefit Is the Bottleneck

Retrieval-augmented generation systems can route difficult queries to deeper context, but batch deployments must allocate a shared token budget across calls whose costs vary by query. We formulate selective escalation as finite-batch allocation for financial document question answering. Each of 150 FinanceBench questions first receives a top-1 retrieval answer. Predictors estimate the adjudication-quality gain and token cost of an optional top-5 call, and the allocator prioritizes calls by predicted gain per token. At the nominal 10% budget, gain-per-token allocation improves adjudication quality over gain-only ranking by 0.034 (95% document-bootstrap confidence interval [0.001, 0.072]) while using 46.6% fewer total tokens than one-pass top-5 retrieval. Additional-call cost is accurately predictable (R-squared 0.93), whereas beneficial escalation remains difficult to rank (area under the receiver operating characteristic curve 0.60). These results show that heterogeneous cost is actionable under tight constraints, while progress across the full budget frontier depends on stronger query-specific benefit estimates.

cs.IR↗

Who Speaks, Who Authorizes, Who Acts? Understanding Trust Boundaries in Multi-Agent LLMs

Multi-agent LLM systems exchange messages before acting through tools, but message delivery, signing authority, and tool-side enforcement are separate trust boundaries. We represent them as communication, authorization, and execution structures, and use the authority cut to characterize the smallest independently compromised signer coalition that can construct a policy-valid false-claim proof under stated assumptions. Controlled paired hidden-state studies reveal four distinctions. First, no fixed-message routing contrast across four models meets the prespecified large-effect criterion; when relays rewrite messages, executed false-state action rates vary by model and route configuration. Second, presented source-identity cues do not uniformly increase protected-action selection across three models. Third, on invalid-proof cases, Nemotron and Llama request protected actions in 52.1% and 54.2% of trials, respectively, yet a deterministic tool gate commits none: request-level and commitment-level safety readouts diverge. Finally, a separate single-model authorization check shows that a false claim with a policy-valid compromised-signer proof can pass the gate, since proof validity is not factual truth. These findings motivate evaluating delivered information, qualified authorization, model proposals, and committed actions separately.

cs.AI↗

Failure-Transparent Agents: Benchmarking Post-Failure Reporting in Tool-Using Language Models

Tool-using agents can fail twice: a required tool can fail, and the agent can then report success without the evidence needed to justify it. Existing benchmarks often entangle this reporting failure with tool selection, recovery, and environment dynamics. We introduce Failure-Transparent Agents (FTA), a controlled benchmark that fixes the failed observation and required evidence state before generation, making post-failure claims directly auditable. FTA contains 100 tasks with deterministic failure traces spanning five failure families, a neutral control, and four user-pressure conditions, and evaluates unsupported claims alongside useful recovery. Across six models, three response policies, and 3,600 human-annotated responses, false-success rates are 22.8% under the baseline policy, 9.3% with a transparency instruction, and 0.8% with a structured evidence contract. Fabricated-detail rates decrease from 28.3% to 14.3% and 0.8%, while useful responses increase from 74.9% to 89.2% and 98.8%, respectively. The tested evidence-contract policy is associated with substantially lower post-failure reporting errors while useful-response rates remain high within this blocked-task benchmark.

cs.AI↗

LeaseGuard: Incumbent-Preserving Admission Control for Privileged LLM Agents

Privileged language-model agents can satisfy a new system task by displacing a healthy incumbent that depends on the same file, process, socket, lock, or capacity allocation. This failure arises because execution privilege determines whether an operation can run, not whether the requester may preempt the current resource owner. We present LeaseGuard, a deterministic admission layer that represents preemption authority through canonical resource leases, incumbent-health checks, effect-aware admission, coexistence limits, safe alternatives, and resource-scoped overrides before adapter execution. We evaluate it on a frozen benchmark of 60 newly authored conflict scenarios with matched controls across two local model families. Relative to a preservation prompt, LeaseGuard reduces unauthorized preemption from 73.3% to 0.0% and increases safe completion by 70.0 percentage points (scenario-clustered 95% CI [60.8, 79.2]). Requested-task success changes by -3.3 points (95% CI [-9.2, 2.5]). The fully evaluated v0.2 broker also rejects a forged incumbent task identity in a hash-linked stress audit. Expiry-only reclamation can still expose a healthy incumbent after missed renewal. The evidence supports incumbent-preserving admission when effects are completely mediated, task ownership is authenticated, and lease expiry reflects incumbent liveness.

cs.CR↗

Residual Community Prototypes Under-Reject Held-Out Malware Families in FCG-MFD

Open-set malware-family recognition must classify known families while rejecting families absent from training. We test whether Louvain-community summaries add rejection information beyond a graph neural network embedding and dimension-matched generic topology. The study uses a deduplicated, conflict-audited FCG-MFD corpus, five held-out families, and three optimization seeds. Community features are residualized against generic topology using known-family training data before nearest-prototype scoring. Residual community does not produce stable held-out-family rejection. Ranking effects reverse across families, the false-positive rate at 95 percent unknown recall worsens for every held-out family, and a validation-fitted threshold rejects only 4.48 percent of unknown samples. Accepted-known macro F1 improves in every family, but with five independent family units the exact two-sided sign-flip p-value is 0.0625, the smallest attainable value. The score remains associated with graph scale, while simple classifier uncertainty performs better on ranking, high-recall rejection, and OSCR. In this GIN/FCG-MFD setting, community-enriched prototypes change known-class geometry without creating a stable unknown margin. Graph open-set evaluations should pair structural features with matched topology controls, operational thresholds, and held-out-family analysis.

cs.CR↗

Partition-Matched Evaluation of Community Features under Distribution Shift in Android Malware Function-Call Graphs

Graph-based Android malware classifiers can lose accuracy under malware-type or family shifts. We test whether mesoscopic organization in function-call graphs provides shift-stable information beyond local degree profiles (LDP), global statistics, lightweight metadata, and size-matched random partitions. Using 15,000 MalNet-Tiny, Common, and Distinct graphs, six Leiden descriptors specified before evaluation, and five optimizer seeds, communities raise Tiny macro F1 from 78.7% to 81.3% but yield 29.4-point source-only Common degradation. One size-matched random partition yields 27.8-point degradation. Across five random partitions, mean degradation is 27.9 points; the 95% two-level bootstrap interval for random minus community degradation is [-4.2, 1.2] points. With metadata, the corresponding difference is -0.2 points with interval [-1.1, 0.6]. Removing modularity raises structure-only Common macro F1 from 51.9% to 53.6%. The tested signature adds IID signal but shows no repeatable shift-stability advantage, demonstrating why mesoscopic graph claims need partition-matched controls and repeated null draws.

cs.CR↗