DataShield is a framework designed to identify risky fine-tuning data for large language models (LLMs) by measuring alignment with consensus safe and unsafe subspaces derived from multiple safety-aligned models. The method enables both sample-level filtering and segment-level masking, reducing attack success rates by 14.6% and 32.3% respectively, while preserving downstream utility.
Why it matters: This work offers a transferable approach to mitigating safety degradation during LLM fine-tuning by filtering risky data without requiring target-model-specific computation.
MARS is a knowledge graph question answering (KGQA) approach that integrates large language models (LLMs) with knowledge graphs without requiring model fine-tuning. It uses a structured retrieval process to iteratively gather relevant information and generate SPARQL queries, adapting the retrieval depth to the question. MARS demonstrates competitive performance on established KGQA benchmarks and is efficient and scalable.
Why it matters: MARS offers a scalable way to improve the reliability of LLMs in knowledge-intensive tasks by grounding answers in explicit, updatable symbolic knowledge without the need for costly fine-tuning.
D-Cut is an adaptive pruning method for speculative decoding in large language models that selects draft tokens jointly across a batch, focusing verification resources on tokens most likely to be accepted. By using cross-request pruning and a runtime cost model, D-Cut adapts to different deployment environments. Experiments show that under high concurrency, D-Cut increases average speedup from 1.26× to 1.65× and achieves up to 3.0× speedup on mixture-of-experts (MoE) models compared to autoregressive decoding.
Why it matters: D-Cut offers a significant advance in efficient large language model inference, particularly under high concurrency, by reducing wasted computation and improving throughput without compromising output quality.
A new preprint compares a GPT-5-based chatbot without scaffolding to a multi-module system (ASuS) that wraps the smaller GPT-4o-mini model in a LangGraph harness for academic supervision. The harnessed system achieved a pooled mean score of 4.08 versus 1.23 for the larger, unscaffolded model across ten independent raters, excelling in reliability, consistency, and other harness-related dimensions. The study's ablation analysis further suggests that the benefits of harness engineering are largely independent of the underlying model size.
Why it matters: This work provides strong evidence that deliberate harness engineering can yield more reliable and consistent AI systems than simply scaling up model size, especially in high-stakes, structured domains.
A new preprint proposes that evaluations of security agents should incorporate economic efficiency, not just task success, by comparing models at fixed cost levels. The study finds that offensive CTF agent performance improves with increased compute resources, while defensive SOC agent success relies more on disciplined tool use and telemetry navigation than on raw reasoning budget. The authors introduce a cost-aware evaluation framework and provide an interactive website with their results.
Why it matters: This work introduces a cost-aware evaluation framework that could reshape how security agents are benchmarked for practical operational use.
Policy & Safety→Official→arXiv Cryptography and Security
A large-scale study of 6,051 Android apps reveals significant inconsistencies between privacy policies and Google Play Data Safety labels, particularly for sensitive data types such as personal information and device identifiers. The research finds that sharing disclosures are less consistent than collection disclosures, and that apps involving persistent monitoring and communication present higher privacy risks. The study introduces a sensitivity-weighted risk score to better assess these misalignments.
Why it matters: This work exposes structural weaknesses in current privacy disclosure mechanisms, highlighting the need for improved verification and transparency in app data practices.
Policy & Safety→Official→arXiv Cryptography and Security
A new preprint introduces MemPoison, a benchmark and analysis framework for evaluating persistent memory attacks on LLM agents. The study covers 1,227 hand-validated attack cases across four attack types and three injection channels, revealing that while write-time defenses can suppress direct attacks, they are ineffective against more complex compositional and context-triggered attacks. The authors propose a taxonomy of attack types and demonstrate structural blind spots in current defense mechanisms, advocating for adaptive, context-sensitive memory defense strategies.
Why it matters: This work exposes critical and previously underappreciated security vulnerabilities in LLM agents with persistent memory, indicating that current defenses are inadequate against sophisticated multi-step attacks.
Researchers introduce T5-CSBoost, an extension of the T5-Sentinel framework that incorporates contrastive style regularization to improve the robustness of AI-generated text detection. The method achieves state-of-the-art results on multiclass source attribution and binary detection benchmarks, and maintains high accuracy under adversarial perturbations up to 90% intensity, including on challenging stress-test suites with unseen models and domains.
Why it matters: This work demonstrates a practical advance in making AI-generated text detectors more resilient to paraphrasing and other adversarial attacks, addressing a key limitation of current systems.
Policy & Safety→Official→arXiv Cryptography and Security
Researchers demonstrate that large language models (LLMs) used for network security log analysis are vulnerable to passive prompt injection, where adversaries embed malicious payloads in log fields that execute when queried. Their LogInject framework achieves up to 88.2% attack success rate across models, and a novel 'Context Stitching' technique evades stateless filters with 76.4% success. Layered defenses reduce attacks by 90.4%, but 8.4% residual vulnerability remains, highlighting the need for defense-in-depth and human oversight.
Why it matters: This research exposes a critical security flaw in LLM-based security operations, showing that untrusted log data can hijack model outputs and undermine threat detection.
Researchers have developed a methodology to automatically extract instruction-centric leakage contracts for major CPU architectures, implemented in the tool malcos. The approach is evaluated on x86 and ARM CPUs, demonstrating that the synthesized contracts are precise and sound with respect to observed microarchitectural leaks.
Why it matters: Automating the creation of leakage contracts can significantly reduce manual effort in securing CPUs against microarchitectural side-channel attacks.
Researchers introduce MCPEvol-Bench, a benchmark designed to evaluate the adaptability of LLM agents to evolving tool interfaces in Model Context Protocol (MCP) servers. By applying 11 mutation operators to 123 MCP servers, they assess 12 state-of-the-art LLMs and find that even advanced models like GPT-5.4 and Claude-Sonnet-4-6 experience significant performance declines—13.7% and 14.4%, respectively—when faced with evolved tool environments. The study also notes increased planning and reasoning errors under these conditions.
Why it matters: This work demonstrates that current LLM agents have notable vulnerabilities when adapting to changing tool interfaces, highlighting a critical challenge for their reliable deployment in dynamic real-world settings.
SAGA is a training-free framework designed to improve agentic text-to-SPARQL generation by systematically incorporating schema constraints. By maintaining a bidirectional type state and filtering out incompatible property candidates, SAGA reduces the search space and prevents semantically invalid queries. The framework achieves the highest F1 scores across all nine benchmark settings on Wikidata and Freebase, and the highest exact-match accuracy on eight of them, while also reducing empty-result queries.
Why it matters: This work systematically addresses type-blind grounding in interactive KBQA agents, leading to more accurate and efficient SPARQL query generation.
Alipay-PIBench is a new benchmark designed to evaluate coding agents on realistic Alipay payment integration tasks. It features nine product-specific projects and 18 task instances, covering both basic functional-completion and advanced risk-aware hardening scenarios. The benchmark uses rubric-based and LLM-assisted assessments to evaluate agent performance. Experiments with six coding-agent models show mean rubric pass rates from 68.58% to 91.37% under the with-skill condition, and access to the payment integration skill improves performance by an average of 10.31 percentage points.
Why it matters: This benchmark enables controlled, realistic evaluation of coding agents on complex, multi-step payment integration tasks, which is important for advancing AI deployment in financial software development.
Researchers present a method for collecting step-level human preference annotations in LLM-based generative agents used in social simulations, resulting in a dataset of 57,000 annotations. By applying supervised finetuning and direct preference optimization to open-weight models using this data, they report consistent improvements in simulation fidelity, coordination, and interaction quality.
Why it matters: This approach offers a scalable way to align agent decision-making with human preferences at a fine-grained level, enhancing the realism and effectiveness of social simulations.
Researchers introduce SportD, a new benchmark that tests vision-language models (VLMs) on soccer decision-making using 478 on-ball events from the 2022 FIFA World Cup. The best-performing VLM selects the optimal action 31.4% of the time, compared to 38.9% for professional players, and models tend to favor safer, less progressive actions. The benchmark quantitatively evaluates models' ability to make strategic decisions in dynamic, real-world scenarios.
Why it matters: SportD highlights a measurable gap between current VLMs and human experts in physical strategic reasoning, providing a new testbed for advancing AI decision-making in complex environments.
MathCoPilot is a human-in-the-loop system designed to facilitate collaboration between mathematicians and AI agents in formal theorem proving. The system integrates an interactive workbench, automated proof orchestration, and topic-driven paper formalization, allowing users to guide and refine proofs interactively. Evaluations using four state-of-the-art LLMs demonstrate strong performance on undergraduate-level problems, but highlight ongoing challenges with more advanced, domain-specific theorems.
Why it matters: This work introduces a novel paradigm for mathematical research, enabling more effective collaboration between humans and AI in formal proof verification and potentially accelerating progress in mathematics.
Researchers present Action QFormer, a query-based interface that reorganizes multimodal information into action-facing representations prior to action generation in vision-language-action (VLA) models. In zero-shot sim-to-real navigation tasks, Action QFormer increases average closed-loop task success from 18.8% to 56.3% and action-generation correctness from 22.5% to 75.5%. The approach reduces destabilization of language-side representations caused by direct action supervision and nearly eliminates out-of-distribution instruction generations.
Why it matters: This work offers a principled method to improve VLA model performance by addressing the trade-off between action supervision and representation stability, advancing beyond reliance on stronger pretrained backbones.
Researchers introduce CEDI, a framework that evaluates vision-language models (MLLMs) through dynamic, multi-turn interactions rather than static benchmarks. When applied to visual hallucinations, CEDI uncovers significantly more hallucinations that better reflect real-world usage, including those that accumulate over long contexts and those triggered by premise rejection. The framework uses a three-party interaction and diverse probing strategies to elicit more ecologically valid evidence of model performance.
Why it matters: CEDI offers a more realistic and systematic method for assessing the reliability of vision-language models, addressing the limitations of static benchmarks.
WrAFT is a modular automated writing evaluation system designed to score and provide feedback on argumentative essays. It achieves state-of-the-art scoring performance (QWK 0.84, RMSE 0.44) on a TOEFL benchmark and receives high human approval ratings for its feedback (over 93%). The system leverages LLMs such as LLaMA-3.3-70B-Instruct, GPT-4o, and Claude 3.7, and is publicly available for use.
Why it matters: WrAFT demonstrates a significant advance in automated essay evaluation by modularizing LLMs for both accurate scoring and high-quality feedback, with practical applications in education.
A new study demonstrates that depth-recurrent transformers with weight-tied cores exhibit per-token fixed-point convergence: the mean KL divergence between successive outputs drops sharply with each loop, reaching near-zero by loop 16. Convergence is non-uniform across tokens, with the median token stabilizing by loop 6, but about 10% requiring up to 8 loops. The authors introduce a simple, training-free early-exit rule that halts processing for each token upon stabilization, achieving the same quality as uniform depth-8 inference while reducing average computation by 38%. This approach outperforms a learned linear router, which fails to reduce computation.
Why it matters: This work reveals a practical, training-free method to reduce inference cost in depth-recurrent transformers by leveraging natural per-token convergence, potentially improving efficiency without sacrificing output quality.