A new framework called AMLGuard is introduced to improve anti-money laundering (AML) in decentralized finance (DeFi) by combining static rule-based analysis with retrieval-augmented large language model (LLM) reasoning. AMLGuard infers high-level transaction semantics to track illicit fund flows, including across blockchains, and was evaluated on 82 real-world laundering cases involving over $1 billion in assets. The system demonstrated high precision and recall in reconstructing illicit fund-flow topologies for both single-chain and cross-chain transactions.
Why it matters: This work represents a notable advance in applying LLMs to financial security, offering improved tools for tracking and analyzing complex illicit cryptocurrency flows.
A new method enables frozen language-model agents to self-evolve their dialogue skills by predicting whether a given response will lead to positive or negative user feedback, using only fixed logged data. This 'future-feedback prediction' approach achieves over 75% accuracy on a proprietary sales-assistant dataset and allows for reproducible, offline optimization of conversational skills without requiring live user interactions.
Why it matters: This work introduces a verifiable offline optimization stage for conversational AI, addressing the challenge of evaluating counterfactual responses and potentially enabling more scalable and reliable self-improvement.
Researchers have identified a critical failure mode in long-context large language models (LLMs) called repetitive copying, where models copy input text into their reasoning traces instead of engaging in productive problem-solving. They introduce GEAR, a reward shaping method that encourages grounding in key evidence and penalizes copying from irrelevant context, leading to consistent improvements of up to +4.6 average points over standard reinforcement learning approaches across multiple benchmarks and model scales.
Why it matters: This work highlights and addresses a pervasive limitation in long-context LLMs, demonstrating that improved evidence grounding can significantly enhance reasoning performance and reduce unproductive copying.
A new verification framework called RECEIPT aims to make LLM-based agent-reported XSS findings trustworthy by enforcing environment isolation, proof-of-concept constraints, role separation, and verdict binding. In evaluations on 95 real-world web applications, RECEIPT discovered 24 previously unknown XSS vulnerabilities within a $20 per-app budget, with 12 acknowledged by maintainers, and recovered labeled CVEs in 36% of known-vulnerability targets, with no false positives reported.
Why it matters: This work addresses the critical problem of reward hacking in LLM-based security agents, providing a deterministic verification method that could make automated vulnerability discovery more reliable and trustworthy.
A new preprint introduces HALLMARK, a benchmark designed to evaluate large language model (LLM) citation verifiers using 2,526 BibTeX entries and 14 types of citation hallucinations. The study finds that the false-positive rate, rather than recall, is the primary factor limiting the practical deployment of citation verification tools, with most LLMs tending to over-flag citations, especially for papers published after their training cutoff.
Why it matters: This benchmark provides a standardized way to diagnose and improve LLM citation verifiers, addressing the risk of fabricated references in AI-generated academic writing.
A new preprint investigates cross-lingual factual inconsistency in large language models (LLMs), where answers shift depending on the prompt language. The researchers evaluate four inference-time interventions on Gemma 3 12B Instruct, including persona prompting, internal representation manipulation, and weight modification. They find that persona prompting—a simple contextual intervention—most effectively balances efficacy, safety, and generalization across languages, outperforming more invasive methods.
Why it matters: This work highlights a practical approach to improving factual consistency in multilingual LLMs, suggesting that simple prompt-based methods can be more robust and transferable than complex interventions.
Researchers present Fusion Embedding, a family of models that extend a frozen vision-language embedding base to include audio without updating its parameters. The first generation uses a 16.4M-parameter connector for audio, while the second adds modality-gated deep adapters (44.2M parameters) that activate only for audio inputs. The models enable zero-shot audio-image retrieval by aligning audio to text, and both generations can be trained in hours on a single GPU.
Why it matters: This work demonstrates a practical method for unifying text, image, video, and audio embeddings in a single space, enabling cross-modal retrieval without requiring paired audio-visual training data.
A new preprint introduces CryptanalysisBench, a benchmark comprising 191 cryptanalysis tasks across six families of cryptographic primitives. Testing five frontier large language models, the study finds that these models can break 65-86% of Tier 1 (known-break) schemes and also produce novel attacks, including a key-recovery attack on SpoC AEAD and the identification of an error in KINDI's security proof. The benchmark is released to track AI progress in cryptanalysis and to stress-test cryptographic schemes.
Why it matters: This work highlights the rapidly advancing capability of LLMs to perform automated cryptanalysis, raising both opportunities and concerns for digital security.
A new framework reformulates the task of mapping automotive CVEs to the Auto-ISAC Automotive Threat Matrix (ATM) as a selective automation problem. Using hierarchical in-context learning and a calibrated meta-model, the system assigns mappings to AUTO, REVIEW, or HOLD tiers based on confidence. In evaluation, the AUTO tier achieved a precision of 0.878 in high-confidence mode, substantially outperforming a zero-shot GPT-5.2 baseline.
Why it matters: This approach advances automated vulnerability mapping in automotive cybersecurity by enabling selective automation and reducing the risk of misclassification in safety-critical contexts.
A systematic measurement study assesses the effectiveness of current web bot defenses against LLM-based browser agents and commercial captcha-solving services. The study finds that challenge-based defenses like hCaptcha and reCaptcha v2 are broadly ineffective, as they can be bypassed by both commercial solvers and LLM agents with solver modules. Non-interactive defenses such as reCaptcha v3 show stronger resistance, but this is due to checks on execution-environment authenticity rather than agent behavior. The results indicate that the true security boundary for these defenses is at the environment layer.
Why it matters: This work highlights a fundamental shift in web security, showing that as LLM agents become more capable, traditional bot defenses are increasingly ineffective and future protections must focus on environment authenticity.
Researchers model Chain-of-Thought reasoning in language models as a switching dynamical system, revealing that reasoning fine-tuning globally reorganizes latent dynamics rather than merely improving token-level competence. The study finds that the recovered latent policy states show functional specialization aligned with reasoning stages, and causal interventions demonstrate their functional significance. Additionally, SDS-guided pruning of failure-prone reasoning prefixes outperforms self-consistency in 11 of 12 settings, with gains up to 12.5 percentage points.
Why it matters: This work introduces a novel mechanistic framework for analyzing and controlling reasoning in language models, with practical implications for enhancing reasoning performance through process-level interventions.
A new preprint stress-tests leading medical AI models in open-ended clinical conversations with missing information, revealing that the choice of LLM judge significantly alters apparent safety ratings. The study finds only moderate agreement between judges and identifies a same-provider bias, where models appear safer when evaluated by their own provider's judge. LLM judges are also more lenient than human clinicians, indicating that observed safety gaps are due to calibration differences rather than knowledge deficits.
Why it matters: This work highlights that current evaluation practices for medical AI may overstate safety due to evaluator bias, underscoring the need for independent, human-anchored assessments before clinical deployment.
A new preprint proposes using semantic primes from the Natural Semantic Metalanguage (NSM) as foundational elements to explain and control emotions in large language models (LLMs). Experiments on four instruction-tuned LLMs demonstrate that NSM primes are more recoverable, controllable, and selective than appraisal-based directions, with interventions on NSM primes controlling emotion about three times as strongly and twice as selectively. The study suggests that NSM primes may serve as more effective explanatory primitives for emotion in LLMs than existing approaches.
Why it matters: This work introduces a potentially more interpretable and effective framework for understanding and manipulating emotions in LLMs, which could impact model alignment and safety.
A new framework called DWM (Decomposed World Model) is proposed to separate latent world model transitions into action-driven and action-invariant components. This approach enables the model to distinguish between changes caused by agent actions and those arising from intrinsic environment dynamics, such as gravity or inertia. In experiments on modified control benchmarks with persistent world effects, DWM achieved a mean absolute improvement of 13.1% in planning success compared to strong baselines.
Why it matters: This decomposition could lead to more transferable and interpretable world models, improving model-based control in environments with complex dynamics.
A new framework called TARA (Type-Aware Repair Allocation) is introduced for optimizing prompts in text-to-image generation. Unlike previous methods, TARA routes each failed proposition in a prompt to a type-specific repair operator, enabling more precise corrections without retraining the generator. Experiments show TARA achieves the highest semantic accuracy across eight benchmark-generator combinations, outperforming VisualPrompter by 5.6 and 2.6 points on DSG and TIFA benchmarks, respectively, and operates faster at 16.0 seconds per prompt.
Why it matters: TARA offers a novel, training-free approach to improving the reliability and accuracy of text-to-image generation, addressing heterogeneous prompt failures more effectively than prior methods.
Researchers have identified a novel attack surface in graph foundation models: the alignment layer that maps diverse graph inputs into a shared representation. They demonstrate that targeted perturbations in this representation space can severely degrade model performance, with some models—such as OpenGraph—being particularly vulnerable. Additionally, input-space attacks that modify edges, features, or text can eliminate at least half of correct predictions in three out of six tested models.
Why it matters: This work exposes a previously unexamined vulnerability in graph foundation models that can be exploited at inference time without access to training data, raising new security concerns for real-world deployments.
A collaboration between the European Commission's Directorate-General for Translation and the European Master's in Translation network has localized the MMLU dataset into 11 European languages. The project not only aims to create a more inclusive benchmark for evaluating large language models (LLMs), but also provides master's students with authentic, project-based professional training in translation and related skills.
Why it matters: This initiative addresses the lack of multilingual evaluation benchmarks for LLMs in European languages and demonstrates a novel approach to dataset creation through educational partnerships.
Researchers have introduced SkillSight, a training-free retrieval framework designed to improve skill selection for large language model (LLM) agents by calibrating shared descriptive background in skill libraries. SkillSight addresses the bias introduced by recurring descriptive patterns in skill descriptions, leading to more accurate retrieval. Experiments show that SkillSight improves Recall@10 by up to 20.21 percentage points and is up to 1,248 times faster than dense+reranker baselines.
Why it matters: SkillSight offers a significant advance in efficient and accurate skill retrieval for LLM agents, which is essential as skill libraries grow larger and more complex.
Researchers introduce Black-Mamba, a test-time adaptive forecasting architecture that uses evidence-gated state tracking to distinguish persistent distribution shifts from transient noise. The model updates its internal memory only when accumulated surprisal signals a regime change, reducing unnecessary updates while maintaining competitive or improved predictive performance on non-stationary forecasting benchmarks. This approach is inspired by biological mechanisms and is supported by mathematical analysis.
Why it matters: This work presents a novel, principled method for efficient and robust online adaptation in AI systems facing non-stationary environments.
Researchers introduce SAGE (ScAffolded Generative models for Explanation), a neuro-symbolic framework that integrates language models with cognitive models to address pragmatic reasoning. SAGE decomposes pragmatic tasks into proposers (generating alternatives), evaluators (assessing alternatives), and selectors (implementing rule-based decisions), and is tested on referential expression generation, M-implicatures, and Gricean implicatures. The framework achieved high accuracy and often outperformed baselines, but analysis revealed that language model proposers excel at generating alternatives, while evaluators are better at intuitive rather than formal judgments.
Why it matters: This work represents a notable advance in computational pragmatics by providing a flexible, interpretable approach that leverages both cognitive modeling and generative language models to improve context-sensitive language understanding and production in AI.