A new study evaluates LLM performance on statics problems using a model distillation approach. LLMs perform well on text-only problems but accuracy drops when diagrams and multi-step reasoning are introduced. The decline is primarily due to difficulties in multi-step reasoning, not image recognition limits.

A study finds that post-training for helpfulness (SFT and RL) significantly degrades animal compassion values instilled during mid-training, while coding post-training better preserves them. Helpfulness training also causes a large drop in English general moral reasoning but not cross-lingually, whereas the compassion degradation transfers consistently across languages. This suggests mid-trained values are encoded more deeply and cross-lingually than reasoning improvements from domain-specific post-training. The paper recommends coding post-training for value-preserving model development.

A new benchmark, Know2Guess, aims to evaluate LLMs' ability to distinguish between knowledge-based answering and guessing, considering data contamination. It includes 1,200 items across five domains and tests models like FLAN-T5, Qwen2.5-Instruct, and Llama-3-Instruct. Qwen2.5-3B-Instruct shows best reliability but still has calibration issues.

HierBias is a novel hierarchical context-conditioned media bias detector that formally models document context for bias prediction, theoretically proving reduced Bayes error and improving sample efficiency via multi-task learning. It achieves 0.853 F1 and 0.723 MCC on BABE and BASIL, surpassing state-of-the-art.

This research proposes a neural network method incorporating topological data analysis to improve flood detection accuracy and interpretability. Using the SEN12-FLOOD dataset, topological features extracted from images are integrated into neural networks, demonstrating that topological descriptors carry meaningful flood signals independently and enhance the robustness and interpretability of existing networks.

Modern machine learning systems have evolved into complex socio-technical architectures that actively mediate human opportunity. The field of algorithmic fairness addresses how models optimized for predictive accuracy can systematically disadvantage marginalized groups. This thesis (arXiv:2606.26200) identifies two fundamental limitations: reliance on deterministic point estimates for auditing and treatment of individuals as isolated entities devoid of structural context.

A new FHPLF model integrates hash learning with federated learning, using binary gradient matrices, projected Hamming distance, and a privacy-enhanced upload strategy to improve accuracy, efficiency, and privacy.

This paper introduces Clue-Guided Group Discovery (CGGD), a paradigm that progressively recovers money laundering groups from initial clues through analyst interaction. The proposed Clue2Group framework constructs a compact local investigation context, uses a multi-semantic local-temporal GNN to estimate risk fields, and integrates evidence to recover criminal groups. Experiments on large-scale AML benchmarks demonstrate its practical potential for real-world investigations.

This paper challenges the assumption that setting LLM-as-judge temperature to 0 ensures deterministic grading, showing flips occur due to default temperature and residual non-reproducibility even under greedy decoding.

KG-TRACE is a novel neuro-symbolic framework that integrates the WHO mutation knowledge graph as a structured biological constraint into a neural genomic model, dynamically weighting neural evidence against symbolic biological knowledge via a learned epistemic trust gate. Evaluated on the CRyPTIC M. tuberculosis cohort, it achieves an AUROC of 0.9760 for isoniazid and introduces the Biological Grounding Ratio (BGR) metric to quantify alignment between neural attributions and established biology.

This paper provides a comprehensive review of Neural Architecture Search (NAS) methods applied to Generative Adversarial Networks (GANs), comparing search strategies, evaluation metrics, and performance outcomes. It highlights the superiority of evolutionary and gradient-based methods in certain contexts, the need for robust evaluation metrics beyond IS and FID, and the importance of diverse datasets.

This study reframes phototaxis in unicellular algae as an information-driven sensorimotor process, linking a POMDP with biochemical reaction dynamics via CRN-ODEs. Using inverse reinforcement learning on 30 Chlamydomonas trajectories, it infers behavioral objectives and shows that run–tumble alternation emerges as an information-acquisition strategy, demonstrating how intracellular biochemical networks support adaptive information-seeking behavior.

A GPU-native population optimizer, χ-sao (Convergence-Halt-Invert-Stick-And-Oscillate), exploits a deliberate convergence-anticonvergence oscillation cycle to escape local traps while freezing confirmed modes. On all 42 functions of the Simon Fraser University optimization benchmark suite across dimensions d ∈ {2,4,8,16,32,64}, χ-sao achieves 100% mode recovery where all CPU baselines collapse at d ≥ 8 on the hardest multimodal functions, with speedups up to 34× over basin-hopping on Michalewicz d=64 and up to 39× on Rotated Hyper-Ellipsoid d=64. Under substantial likelihood noise (σ_noise up to 1.0), mode detection remains 100% reliable. The algorithm is available as an open-source Python package on PyPI.

Researchers propose an attention-based, physics-guided convolutional neural network as a surrogate model to predict microstructural evolution in systems governed by the Cahn-Hilliard equation. The model accurately forecasts phase separation in binary mixtures over long times, preserves composition, and aligns with the Lifshitz-Slyozov domain-growth law.

A classical intuition holds that verifying a solution is easier than producing one. For today's coding agents, this intuition is inverted: generating complex candidate solutions is no longer difficult, but reliably verifying them has become the harder problem. Every verifier is a proxy for human intent, never the intent itself, leading to a twofold difficulty: underspecified intent and optimization widening the proxy–intent gap. The paper characterizes verification signals along scalability, faithfulness, and robustness, and studies four reward constructions. Experiments show targeted verification design can suppress reward hacking and improve task completion quality, with the core observation that no fixed reward function remains effective as policy capability grows; verification must co-evolve with the generator.

COrigami is an end-to-end AI-driven pipeline that generates crease patterns from natural language, satisfying strict flat-foldability constraints and visual aesthetics. It assists human artists by generating structural starting points through steps including semantic stick figure generation, base packing, crease pattern solving, shaping, and reinforcement learning with an autonomous aesthetic evaluation loop.

This paper proposes a governance model for autonomous AI agents that does not monitor their reasoning but requires independently attested evidence at the point of high-risk actions. The agent retains autonomy over planning and reasoning, but execution of designated high-risk actions is conditional on preconditions attested by separate authoritative sources, cryptographically bound to a declared intent, and evaluated by a deterministic policy. Decisions are recorded in a tamper-evident log. A proof-of-concept implementation is presented with examples from software deployment and clinical prescribing.

Researchers propose DD-Elo, a skill assessment framework using drift diffusion model and move-level data, achieving faster adaptation than Elo while maintaining bounded deviation.

This study develops a provenance-aware, knowledge-graph-based multi-agent framework that integrates Reddit posts, WebMD reviews, and FDA adverse event records for nine antidepressants, achieving high entity recognition accuracy and revealing that patient-generated data provide partly independent safety signals, with community sources often preceding regulatory reports.

This paper introduces an LLM-powered comparative pipeline for large-scale governance discourse analysis of AI agent protocols. It validates the pipeline on two contrasting standards: ERC-8004 (permissionless, on-chain) and Google A2A (corporate-led). Analyzing 4,323 governance participation records, it finds that while governance form influences substantive focus, both regimes exhibit comparable participation inequality and community fragmentation. However, discourse alignment is denser in the permissionless setting, suggesting open governance may foster greater thematic convergence.

AlgoEvolve is an LLM-driven evolutionary framework that generates, evaluates, and iteratively improves executable trading strategies. Across multiple experiments, the system exhibits emergent regime-adaptive strategy logic and introduces a meta-evolutionary outer loop to evolve prompts, improving search heuristics. The results demonstrate that LLM-based semantic evolution provides a viable approach for continual program synthesis in complex environments.

This paper shows that refusal in instruction-tuned chat models is gated by a compliant persona direction. Interventions on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct demonstrate that steering compliant persona suppresses refusal (e.g., Llama's refusal rate drops from 97% to 2%), and refusal direction only partially restores refusal in late layers. The findings indicate refusal is expressed downstream of persona computation.

When a benchmark's accuracy saturates, it is often retired. This paper shows that this approach misses six other key dimensions: construct validity issues, out-of-distribution generalizability, efficiency, reliability, model versus scaffold importance, and human-agent collaboration uplift. Using CORE-Bench Hard, they surface construct validity threats, introduce an improved benchmark v1.1 and OOD suite, and find that the benchmark remains useful for measuring efficiency, reliability, and performance. A small-scale experiment shows human-agent collaboration yields about a 2x speedup.

Researchers propose a method to detect and control sycophancy in language models using cascading linear features. Their approach uses iterative data generation to isolate features that scale linearly with behavior, enabling better disentanglement. The discovered features form linearly separable subspaces, allowing for detection and steering away from sycophancy, outperforming baseline methods with lower computational cost.

LogiGate is a zero-trust middleware architecture that shifts identity validation and legal liability to the requester device's hardware enclave, using ephemeral sandboxes and forced reset triggers to ensure compliance and data privacy in AI processing.

I built hackmyclaw.com, where anyone could email my AI assistant Fiu and try to make it leak a secrets.env file. After reaching HN front page, over 2,000 people sent more than 6,000 emails trying to break it. The secrets never leaked.

This article discusses the necessity of guardrails for offensive AI agents, explaining why mere instructions are insufficient to ensure safety and prevent misuse.

Llamatik Code is a paid AI coding assistant plugin for JetBrains IDEs, emphasizing local-first processing for code privacy and offline capability.

A new bipartisan nonprofit, RAISE US, launches with over $500 million to help American workers adapt to AI-driven job changes through education and training, starting in four states and partnering with major companies like Amazon and Microsoft.

The author, after assembling a list of AI-generated children's books, purchased a bestseller and found disturbing body horror imagery, questioning the quality of current AI content.