A survey of LLM agent self-security threats and mitigations alongside their applications in the cybersecurity lifecycle, introducing a synergy concept and empowerment framework.
hub
and Lewis, Mike , editor =
21 Pith papers cite this work, alongside 152 external citations. Polarity classification is still indexing.
21
Pith papers citing it
152
external citations · Crossref
hub tools
citation-role summary
background 1
dataset 1
citation-polarity summary
representative citing papers
SIOP enables turn-level credit assignment in LLM agents via semantic clustering of final answers as latent outcomes, improving performance on reasoning benchmarks without verifiers.
AI Overviews and Gemini retrieve substantially different sources than traditional Google search (Jaccard similarity <0.2), favor Google-owned content, appear for 51.5% of queries especially controversial ones, and are less consistent across repeated or slightly edited queries.
GDP-RAG targets only information deltas in multi-hop RAG through preliminary grounding, gap-conditioned prompts, and skeletal trajectories, reaching 60.63% accuracy at 0.51 cost-of-pass on HotpotQA, 2WikiMultiHopQA, and MuSiQue.
Controlled experiments show implicit multi-hop reasoning in LLMs requires prior exposure to compositional contexts during pretraining and does not transfer to unexposed individuals.
RASER routers built on one-shot RAG features selectively escalate retrieval, matching SOTA F1 scores on multi-hop QA while using 41-49% of the tokens required by always-prune across six LLMs and three benchmarks.
K-BrowseComp is a new Korean web-browsing agent benchmark where frontier LLMs score 30-46% and Korean LLMs score 0-10% on the verified subset.
Harness-1 uses a state-externalizing harness for RL-trained search agents and reports 0.730 average curated recall, outperforming the next open subagent by 11.4 points.
EAPO learns selective tool use in agentic RL via tool-free trajectories, difficulty-aware reward shaping, and confidence-aware token reweighting, improving accuracy while cutting tool calls versus GRPO on nine reasoning benchmarks.
Dynamic Gradient Gating monitors lm_head gradient norms to safely reuse rollout batches in RLVR, achieving up to 2.93x sample efficiency and 2.14x wall-clock speedup across math, ALFWorld, WebShop, and QA tasks.
ICRL uses joint RL training of solver and critic with distribution-calibration re-weighting and role-wise advantage estimation to internalize critique into unassisted LLM performance, yielding 6.4-point gains on agentic tasks and 7.0 on math reasoning with Qwen3 models.
Synthetic data improves models only in information-open generation-training loops with external signals, and coarser signals like binary correctness enable better generalization by converging to the most information-efficient component.
Verbal-R3 uses a verbal reranker to generate analytic narratives that guide retrieval and reasoning in LLMs, achieving SOTA results on complex QA benchmarks.
Power-law data distributions outperform uniform ones for compositional reasoning by creating asymmetry that lets frequent skill compositions scaffold rare ones with less data.
ARPO adds entropy-based adaptive rollouts and stepwise advantage attribution to RL for LLM agents, outperforming prior trajectory-level methods on 13 benchmarks with half the tool budget.
TASR provides a training-free predicate that stops iterative retrieval on repeated normalized answers plus calibrated logit margin above 0.25, retaining 94.8% of fixed-k=5 F1 at 62.6% of the calls across 32 configurations.
ReCal introduces hierarchical reward decomposition and distribution-aware optimization to address ambiguous credit assignment and optimization bias in RL-based LLM routing.
ANCHOR uses hierarchical factor construction and causal Bayesian networks to reduce unknown predictions and improve reliability of LLM-based probability inference over prior Naive Bayes approaches.
citing papers explorer
-
LLM agents security duality: a comprehensive survey of self-security and empowered cybersecurity
A survey of LLM agent self-security threats and mitigations alongside their applications in the cybersecurity lifecycle, introducing a synergy concept and empowerment framework.
-
Self-Induced Outcome Potential: Turn-Level Credit Assignment for Agents without Verifiers
SIOP enables turn-level credit assignment in LLM agents via semantic clustering of final answers as latent outcomes, improving performance on reasoning benchmarks without verifiers.
-
How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews
AI Overviews and Gemini retrieve substantially different sources than traditional Google search (Jaccard similarity <0.2), favor Google-owned content, appear for 51.5% of queries especially controversial ones, and are less consistent across repeated or slightly edited queries.
-
Only Ask What You Don't Know: Grounded Delta Planning for Efficient Multi-step RAG
GDP-RAG targets only information deltas in multi-hop RAG through preliminary grounding, gap-conditioned prompts, and skeletal trajectories, reaching 60.63% accuracy at 0.51 cost-of-pass on HotpotQA, 2WikiMultiHopQA, and MuSiQue.
-
Multi-Hop Knowledge Composition is Bound by Pretraining Exposure
Controlled experiments show implicit multi-hop reasoning in LLMs requires prior exposure to compositional contexts during pretraining and does not transfer to unexposed individuals.
-
RASER: Recoverability-Aware Selective Escalation Router for Multi-Hop Question Answering
RASER routers built on one-shot RAG features selectively escalate retrieval, matching SOTA F1 scores on multi-hop QA while using 41-49% of the tokens required by always-prune across six LLMs and three benchmarks.
-
K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts
K-BrowseComp is a new Korean web-browsing agent benchmark where frontier LLMs score 30-46% and Korean LLMs score 0-10% on the verified subset.
-
Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses
Harness-1 uses a state-externalizing harness for RL-trained search agents and reports 0.730 average curated recall, outperforming the next open subagent by 11.4 points.
-
Learning When Not to Act: Mitigating Tool Abuse in Agentic Reinforcement Learning
EAPO learns selective tool use in agentic RL via tool-free trajectories, difficulty-aware reward shaping, and confidence-aware token reweighting, improving accuracy while cutting tool calls versus GRPO on nine reasoning benchmarks.
-
When to Stop Reusing: Dynamic Gradient Gating for Sample-Efficient RLVR
Dynamic Gradient Gating monitors lm_head gradient norms to safely reuse rollout batches in RLVR, achieving up to 2.93x sample efficiency and 2.14x wall-clock speedup across math, ALFWorld, WebShop, and QA tasks.
-
ICRL: Learning to Internalize Self-Critique with Reinforcement Learning
ICRL uses joint RL training of solver and critic with distribution-calibration re-weighting and role-wise advantage estimation to internalize critique into unassisted LLM performance, yielding 6.4-point gains on agentic tasks and 7.0 on math reasoning with Qwen3 models.
-
An Information-Theoretic Criterion for Efficient Data Synthesis
Synthetic data improves models only in information-open generation-training loops with external signals, and coarser signals like binary correctness enable better generalization by converging to the most information-efficient component.
-
Verbal-R3: Verbal Reranker as the Missing Bridge between Retrieval and Reasoning
Verbal-R3 uses a verbal reranker to generate analytic narratives that guide retrieval and reasoning in LLMs, achieving SOTA results on complex QA benchmarks.
-
The Power of Power Law: Asymmetry Enables Compositional Reasoning
Power-law data distributions outperform uniform ones for compositional reasoning by creating asymmetry that lets frequent skill compositions scaffold rare ones with less data.
-
Agentic Reinforced Policy Optimization
ARPO adds entropy-based adaptive rollouts and stepwise advantage attribution to RL for LLM agents, outperforming prior trajectory-level methods on 13 benchmarks with half the tool budget.
-
TASR: Training-Free Adaptive Stopping for Iterative Retrieval
TASR provides a training-free predicate that stops iterative retrieval on repeated normalized answers plus calibrated logit margin above 0.25, retaining 94.8% of fixed-k=5 F1 at 62.6% of the calls across 32 configurations.
-
ReCal: Reward Calibration for RL-based LLM Routing
ReCal introduces hierarchical reward decomposition and distribution-aware optimization to address ambiguous credit assignment and optimization bias in RL-based LLM routing.
-
ANCHOR: Abductive Network Construction with Hierarchical Orchestration for Reliable Probability Inference in Large Language Models
ANCHOR uses hierarchical factor construction and causal Bayesian networks to reduce unknown predictions and improve reliability of LLM-based probability inference over prior Naive Bayes approaches.
- LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents
- Skill-RAG: Failure-State-Aware Retrieval Augmentation via Hidden-State Probing and Skill Routing
- Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers