A new preprint demonstrates that log-likelihood-based (perplexity) evaluation can significantly underestimate the performance gap between teacher and distilled student models in autoregressive generation tasks. The authors introduce GenDistill, a multi-stage pipeline for distilling Transformers into efficient Hybrid Kimi Delta Attention models, and systematically analyze the impact of various design choices. They find that dataset selection, completion-only masking, and freezing attention layers are particularly important for maintaining generation quality in distilled models.
Why it matters: This work exposes limitations of perplexity-based evaluation in model distillation and offers practical insights for building efficient, high-quality generative models.
Researchers investigated the Knobe effect—a moral bias in intentionality judgments—in finetuned large language models (LLMs). Using Layer-Patching analysis on three open-weights models, they found that the bias is acquired during finetuning and can be localized to specific layers. Importantly, patching activations from the pretrained model into these critical layers eliminated the bias without retraining the entire model.
Why it matters: This work demonstrates that social biases in LLMs can be interpreted, localized, and mitigated through targeted interventions, suggesting a practical path to safer models without full retraining.
Policy & Safety→Official→arXiv Cryptography and Security
Researchers introduce CPPIA, a novel code-level property inference attack that allows adversaries to embed property information into secret samples during model training. By leveraging poisoned code from platforms like GitHub or coding agents such as Codex, attackers can later query the trained model to extract global properties of the private training data. The attack achieves 100% accuracy, does not degrade model performance, is computationally efficient, and bypasses existing defenses across multiple datasets and architectures.
Why it matters: CPPIA reveals a new and effective privacy threat in machine learning pipelines that depend on third-party code, highlighting the risk of sensitive data leakage even when standard defenses are in place.
A new compiler, Recifhe, introduces multi-level optimization for fully homomorphic encryption (FHE) programs by operating at both the ciphertext and polynomial levels. Unlike previous compilers that optimize only at the ciphertext level, Recifhe eliminates redundant polynomial computations across ciphertext operations, resulting in a reported 1.25x speedup over ciphertext-level-only optimization.
Why it matters: This work advances the efficiency of privacy-preserving computation on encrypted data, which is important for secure cloud-based services.
A new preprint introduces ActiveVision, a benchmark designed to test whether multimodal large language models (MLLMs) can perform active visual observation—redirecting their 'gaze' based on intermediate reasoning, rather than relying on static images. Leading models such as GPT-5.5 and Claude Fable 5 scored only 10.6% and 3.5% respectively, compared to a human average of 96.1%. The findings suggest that current MLLMs lack robust active visual perception, even when allowed to write and execute their own vision code.
Why it matters: This work reveals a fundamental limitation in current MLLMs, highlighting the need for new architectures that integrate perception and reasoning in a closed loop.
Policy & Safety→Official→arXiv Computation and Language
Researchers present Latent Fusion Jailbreak (LFJ), a white-box attack that manipulates the internal hidden states of large language models (LLMs) by interpolating between harmful and benign prompts. LFJ achieves a macro-averaged attack success rate of 94.13% across five open-weight models, but a tailored adversarial training defense reduces this rate to 12.37%. The study also explores the impact of different pairing and sampling strategies on attack effectiveness.
Why it matters: This work reveals a novel vulnerability in safety-aligned LLMs via internal state manipulation and demonstrates a potential defense, underscoring the evolving challenges in AI safety.
A new preprint introduces DECODEM, a benchmark dataset designed to evaluate automated extraction of corporate governance variables from organizational documents such as charters and bylaws. The authors use high-quality human annotations to assess the performance of various large language model (LLM) pipelines on document-level binary classification tasks. Results indicate that automated extraction is feasible and achieves high accuracy for many governance provisions, though performance varies across variables. The study also finds that pipeline design can partly compensate for differences in model capability.
Why it matters: This work provides a standardized benchmark and systematic evaluation, showing that current LLMs can accurately extract legally relevant information from complex corporate documents, potentially reducing the need for manual coding in legal research.
Policy & Safety→Official→arXiv Cryptography and Security
A new preprint reveals that using humor as an indirect refusal mechanism in large language models (LLMs) can introduce latent safety risks, such as stereotypes and toxicity. The authors introduce HumorSafe, a framework for evaluating these risks, and HumorPIA, a prompt injection attack that exploits humor-based defenses to covertly increase toxicity while maintaining a high apparent safety rate.
Why it matters: This research exposes a previously overlooked vulnerability in LLM safety mechanisms, showing that humor-based defenses can covertly propagate harmful content and evade current detection methods.
BayesPO introduces a probabilistic framework for prompt optimization by casting it as Bayesian posterior sampling over discrete tokens, leveraging gradient-guided MCMC with parallel tempering. Experiments with Qwen2.5 models demonstrate that BayesPO can discover semantically meaningful prompts and improve accuracy on instruction-induction tasks, though the approach is computationally intensive and susceptible to overfitting on small datasets.
Why it matters: This work advances prompt optimization by providing a principled, probabilistic alternative to heuristic search methods, potentially leading to more robust and effective prompt discovery for large language models.
Researchers introduce Neural Cryptographic Services (NCS), a security governance framework that cryptographically authorizes each action taken by LLM-based agents. NCS interposes a deterministic, symbolic controller between agents and privileged tools, requiring each tool invocation to match a signed, hash-chained instruction. Experimental results show that NCS reduces prompt injection attack success rates to near zero while maintaining utility on legitimate workflows.
Why it matters: This work proposes a novel cryptographic approach to mitigating prompt injection attacks in autonomous AI agents, potentially enabling safer deployment in sensitive environments.
Researchers introduce Lucid, a black-box adversarial framework that targets multimodal AI agents by crafting imperceptible perturbations to images, compromising their long-term memory pipelines. Lucid enables two attack modes—memory poisoning and memory injection—achieving 61.6% and 58.4% attack success rates, respectively, across five black-box memory architectures, including commercial systems. The attacks require no access to the target model or text channel, operating solely through manipulated visual inputs.
Why it matters: This work exposes a critical vulnerability in multimodal AI agents' reliance on visual data for persistent memory, highlighting the risk of adversarial manipulation through images alone.
Policy & Safety→Reported→arXiv Cryptography and Security
Researchers have shown that speech recognition models can be compromised using ordinary natural sounds as backdoor triggers. Their experiments demonstrate that with only 5% of training data poisoned, these attacks achieve nearly 100% success while leaving model performance on benign inputs unaffected. The use of common sounds makes the attacks stealthy and difficult to detect.
Why it matters: This finding exposes a significant security vulnerability in speech recognition systems, as undetectable backdoors could be triggered by everyday sounds.
A new benchmark, HypoArena, has been introduced to evaluate large language models (LLMs) on their ability to autonomously construct hypothesis spaces from inconclusive evidence across six scientific and analytical domains. The study presents Prospective Hypothesis Discovery (PHD) as a novel evaluation paradigm, and experiments on 15 advanced LLMs reveal clear stratification in capabilities and model-dependent effects of structured analytical skills.
Why it matters: This work establishes a new standard for assessing LLMs' open-ended reasoning before conclusions are drawn, highlighting a distinct and underexplored capability in AI systems.
AV-JEPA is a multimodal extension of LeJEPA for audio-visual self-supervised learning, employing an early-fusion Vision Transformer and modality dropout. The architecture achieves competitive results on VGGSound (57.1% top-1) and AudioSet (32.7 mAP), and enables zero-shot audio-video retrieval without requiring decoders or contrastive negatives.
Why it matters: AV-JEPA demonstrates a streamlined approach to multimodal self-supervised learning, achieving strong performance on audio-visual tasks with a simplified architecture.
A preprint systematically evaluates five large language models (LLMs)—GPT-4 Turbo, Claude 3 Opus, Gemini 1.5 Pro, Llama 3 70B, and FinGPT—for technical market analysis tasks. The study finds that GPT-4 Turbo achieves the highest annualized return and Sharpe ratio among general-purpose models, while FinGPT offers competitive risk-adjusted performance due to domain-specific fine-tuning. Both models outperform a passive S&P 500 benchmark in simulated backtesting, but all exhibit notable failure modes such as numerical hallucination and context-window limitations.
Why it matters: This work provides a rigorous benchmark for LLMs in financial trading, highlighting both their potential and critical limitations for real-world use.
A new approach combines a closed-loop evolutionary algorithm with a large language model (LLM) to automate the design of physics-informed neural networks (PINNs). The system iteratively generates and evaluates complete PINN configurations, using training outcomes to inform subsequent generations. In experiments on a 1D multiscale wave equation, the best configuration emerged in the final generation, achieving up to a 95.38% reduction in mean-squared error compared to the initial population. The results demonstrate the feasibility of this method for automated PINN design.
Why it matters: Automating PINN design could accelerate progress in scientific computing by reducing the manual effort required to optimize neural network architectures for complex physical systems.
A new autoscaling framework for serverless environments combines graph-based dependency analysis, short-term workload forecasting using multiple neural models (MLP, LSTM, CNN), and cost-aware scaling control. The approach uses a probabilistic ensemble of predictors, achieving 99.88% prediction accuracy and reducing infrastructure costs while maintaining performance targets in experiments with real workload traces. The framework also incorporates cold-start awareness and evaluates performance across multiple cloud pricing models.
Why it matters: This research demonstrates a robust and practical advance in serverless autoscaling, addressing key challenges of workload prediction, cost efficiency, and dependency management in cloud applications.
A new method called Recursive Harness Self-Improvement (RHI) iteratively refines prompt-level harness specifications using pairwise feedback over revision history. Tested on 30 synthetic machine learning tasks, RHI substantially increases the performance ceiling of low-reasoning-effort agents, even surpassing the results of maximum-reasoning-effort settings, while reducing inference costs by up to 60%. The improvements are attributed to better task-specific context management and more effective inter-agent information flow, rather than simply longer reasoning traces.
Why it matters: RHI demonstrates a practical, lightweight approach for continually improving agent harnesses, enabling cost-efficient performance gains in model-harness co-evolution.
Researchers introduce PA-DSL, a method that leverages adjudicated cases to correct noisy human audit labels and then debiases automated classifier labels for statistical analysis. The approach is valid for a wide range of downstream analyses when audit and adjudication probabilities are known. Experiments on synthetic and semi-synthetic data show that PA-DSL maintains nominal coverage and reduces RMSE by 10-17% compared to using only adjudicated labels when human labels are noisy but informative.
Why it matters: PA-DSL offers a principled solution to the widespread problem of noisy human labels in supervised learning, improving the reliability of automated analyses.
Researchers introduce SLAPBench, the first benchmark for evaluating multimodal large language models (MLLMs) on four-finger SLAP fingerprint verification, using data from NIST SD302b with 7,832 pairs. The study finds that prompting strategies determine whether verification collapses, while model capability affects discrimination performance. Claude Opus 4.8 achieves the best binary verification result (FAR=20.2%), and Qwen3-VL-8B achieves perfect separation (AUC=1.000) under similarity scoring, though this may reflect dataset artifacts rather than true capability.
Why it matters: This work establishes the first SLAP-specific MLLM baseline and reveals how prompting and model capability interact in high-stakes biometric verification tasks.