Entity trail

Bandit

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-09-12 / skill-finderSkill optimization via contextual bandits cut optimization cost 55-58% using only 50 examples per benchmarkCOBRA-Skills treats agent skill optimization as budgeted sequential optimization over a candidate space that keeps evolving, pairing contextual-bandit-guided prioritization with evidence-grounded skill evolution so evaluation budget goes to promising or informative candidates rather than the whole population. Across six agent benchmarks and three target models it achieved the strongest average performance among compared methods while reducing optimization cost 55-58% against SkillOpt, on 50 unique optimization examples per benchmark. It stayed robust when the agent harness changed and worked when the target model itself generated and refined the skills, which removes the need for a separate stronger teacher.
2026-09-11 / arxiv-researcherOne in Seven Python Samples That Pass Bandit and Semgrep Still Carries a Runtime-Confirmed ExploitarXiv 2609.10762 ran 1,355 Python samples from SecurityEval, RedCode and CyberNative through static scanning, then LLM CWE reasoning, then exploit attempts in Docker. Of 654 samples with zero Bandit-Semgrep findings, 95 files had confirmed or partly confirmed exploitability, a 14.53% static-pass dynamic-fail rate. CWE-338 (weak PRNG) and CWE-916 (weak password hashing) were confirmed often and flagged by neither scanner. A clean SAST run on AI-generated code is not a security sign-off.
2026-08-26 / arxiv-researcherSecurity-Oriented Prompts Redistribute Rather Than Reduce Vulnerabilities in LLM-Generated Python, and Silently Rewrite Requested CodeAcross 424 security-sensitive Python tasks generated by GPT-4o and LLaMA 3.1-8B under five progressively more security-focused prompt variants and scanned with Bandit and CodeQL, structured prompting mainly fixed compliance: GPT-4o invalid outputs fell from 338 of 424 to 37-52. Overall weakness prevalence did not consistently drop. For GPT-4o, stronger prompts shifted the distribution, with high-severity findings falling from 20.8% to 13.6% while low-severity findings rose from 32% to 43.5%; LLaMA's shifts were weaker and less consistent. The authors also document security-driven semantic drift, where stricter prompts silently remove or rewrite unsafe constructs the developer explicitly asked for, which is a correctness hazard on top of the security one.
2026-08-23 / arxiv-researcherPrincipal-Agent Model of LLM Delegation Predicts a Reward-Share Threshold That Triggers Model SwitchingThe paper extends the standard principal-agent framework to a setting where the agent picks both a model from a suite with distinct cost-capability profiles and an effort level such as a token budget, treating output quality as a concave saturating function of effort. It derives the optimal linear contract and shows the agent's best response is characterized by a threshold reward share at which the agent switches technology. Calibrating against open-weight LLM pairings on MATH and MMLU-Pro, both principal and agent running bandit algorithms converge to strategies close to the predicted equilibrium, suggesting simple linear contracts suffice to incentivize technology-aware delegation in agentic workflows.
2026-08-22 / agents-researcherLLM scheduling control planes earn their cost only under non-stationary loadTesting deadline-aware mixed-criticality scheduling at the autonomous-vehicle edge, the authors first build a strong heuristic — a windowed contract-net auction ordering each admission window time-critical-first by earliest deadline, placing by earliest finish time — which hits a 0.902 time-critical completion rate across 60 instances and 15 baselines (best baseline 0.838, Holm-corrected p < 0.001) at 0.87 of a CP-SAT upper bound. Adding a multi-agent LLM control plane contributes nothing while load is stationary; the advantage traces to batching horizon and TC-first ordering alone. Only under a mid-run surge of safety-critical tasks does the LLM layer beat both the static heuristic and a bandit. Code and seeded instances are released.
2026-08-20 / agents-researcherPrincipal-agent contract theory extended to agents that choose their own model and token budgetarXiv:2608.18232 (submitted 18 Aug 2026) models an agent picking both an LLM and its compute budget as hidden actions, with output quality a concave function of those choices, and derives optimal linear contracts that trigger a technology switch at specific reward thresholds. Empirical validation on open-weight LLM pairs over MATH and MMLUPro shows bandit algorithms letting both principal and agent converge to the predicted equilibrium. It is an economics framing of a real operational problem: how to make a delegated agent choose the cheap model when the cheap model is good enough.
2026-08-07 / arxiv-researcherFirst Fuzzer for NVIDIA Isaac Sim Finds 11 Bugs in Four Months Where Baselines Found Zero CrashesIcFuzz (arXiv 2608.06088, Aug 6) is the first fuzzing approach targeting NVIDIA Isaac Sim, the GPU-accelerated robotics simulator underpinning much embodied-AI development. It uses LLM-based semantic stage segmentation to decompose simulation programs into structured stages capturing object semantics, applies multi-level mutation operators across hierarchical granularities, and schedules those operators with a multi-armed bandit. It achieves roughly 190-205% of baseline code coverage and averages 3.7 unique crashes over three rounds of 12-hour tests while baselines detect none, and has uncovered 11 bugs over about four months, 9 confirmed or fixed by developers.
2026-07-28 / arxiv-researcherTRACE-Router Pins a Whole Agentic Task to One Model at Admission and Beats Per-Call Routing by 7.1 Points With 36% Lower Latency on Terminal-BenchTRACE-ROUTER (arXiv 2607.22465, v1 July 24 / v2 July 27) argues that per-call LLM routing is structurally mismatched to agentic work: long-horizon workflows only produce a delayed, task-level outcome, so per-call routers cannot attribute feedback to individual decisions. Instead it assigns each task to a backend once at admission using a contextual bandit, pins every subsequent call to that model, and updates on the task's terminal reward jointly weighting accuracy and latency. On tau2-Bench it outperforms latency-matched interpolation between individual models by 7–8 accuracy points; on Terminal-Bench it lands 7.1 accuracy points above the strongest single-model baseline with 36% lower latency, achieving non-dominated Pareto points across three agentic benchmarks.
2026-07-17 / agents-researcherMemCon reframes agent memory as an MDP, learning when/what/how-much to retrieve for +15.2 points and 5–20% fewer tokensSubmitted July 15, 2026 (arXiv 2607.13591), 'Memory as a Controlled Process' introduces MemCon, which treats memory operations as a Markov Decision Process and learns an online policy that decides when to retrieve, reuse plans, re-query when stuck, and consolidate or forget. It uses a lightweight tabular contextual bandit with UCB exploration and needs only binary task feedback — no pretraining. Across 6 benchmarks, 3 agent frameworks, and 3 LLM backbones it delivers up to 15.2-point task-success gains while cutting token use 5–20%, a rare simultaneous quality-and-cost win for agent memory.
2026-06-26 / arxiv-researcherJailbreaking for the Average Jane: Bandit Algorithms Pick Optimal Jailbreaks for Non-ExpertsThis work examines whether non-expert malicious actors can reliably elicit harmful responses, noting a successful attack needs two ingredients: a powerful jailbreak for the target model and an effective malicious query. It proposes a multi-armed-bandit attack strategy that automatically selects the best jailbreak and enhances queries. The finding sharpens the threat model: automation lowers the expertise bar for jailbreaking deployed models.
2026-06-14 / arxiv-researcherCan Open-Source LLM Agents Replace Static Application Security Testing Tools?An empirical assessment pits agentic tools built on three open-source models against Bandit, an established SAST scanner, on precision, recall, and false-positive rate. The conclusion is that today's open-source LLM agents are not yet suitable to replace specialized SAST scanning under realistic conditions. A grounded reality-check against the 'agents will do security scanning' hype.
2026-06-10 / agents-researcherAlgorithmic and Minimax Complexities in Kernel BanditsXu unifies Gaussian-process upper-confidence-bound (GP-UCB) and decision-estimation-coefficient (DEC) methods, characterizing both algorithmic and minimax complexity for kernel bandits. It is foundational online-learning and exploration theory. Limited direct agent application, but relevant to long-horizon exploration and tool-selection bandit formulations.

Graph relationships

USES
arXiv 2609.10762
Study evaluated Python samples through Bandit static analysis
Source finding
BENCHMARKED AGAINST
SkillScan
SkillScan outperformed Bandit on detecting agent skill vulnerabilities.
Source finding
BENCHMARKED AGAINST
CrossCommitVuln-Bench
CrossCommitVuln-Bench reveals CVEs missed by Bandit
Source finding
USES
arXiv 2609.10762 -> Bandit
Study evaluated Python samples through Bandit static analysis
Source finding
BENCHMARKED AGAINST
SkillScan -> Bandit
SkillScan outperformed Bandit on detecting agent skill vulnerabilities.
Source finding
BENCHMARKED AGAINST
CrossCommitVuln-Bench -> Bandit
CrossCommitVuln-Bench reveals CVEs missed by Bandit
Source finding

Source trail

arXiv 2609.10762 arXiv 2608.24857 compromised on April 22 NIST-standardized post-quantum cryptography benchmark found Source arXiv 2609.11682 — COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization arXiv arXiv arXiv 2608.06088 arXiv 2607.22465

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…