AINeutralarXiv – CS AI · Jun 257/10
🧠Researchers discovered that language models forget learned rules midway through training despite continued evidence in data—a phenomenon called 'natural ungrokking.' The survival of rules depends predictably on how often they appear in training data, and attempts to restore forgotten rules through data manipulation fail despite successfully destroying them, revealing asymmetric control over model knowledge.
AIBullisharXiv – CS AI · Jun 257/10
🧠Researchers introduce BrReMark, a framework that enhances brain MRI diagnosis by requiring AI models to explicitly mark and verify abnormal regions before reaching conclusions. The approach dramatically improves diagnostic accuracy and reduces false positives by 45.7% on out-of-distribution data, addressing critical trust and hallucination issues in medical AI systems.
AINeutralarXiv – CS AI · Jun 257/10
🧠Researchers introduce Xcientist, a research harness that makes AI scientific reasoning transparent and auditable by externalizing research synthesis into inspectable artifacts. The system addresses 'claim drift'—where AI-generated mechanisms lose evidential grounding—and demonstrates traceable workflows across three scientific domains, suggesting AI scientists should be evaluated on accountability and reproducibility, not just output.
AIBearisharXiv – CS AI · Jun 237/10
🧠A new research framework reveals that large language models exhibit inconsistent behavior across structurally equivalent decision environments, demonstrating significant portability losses when behavioral patterns learned in one setting are applied to another. The findings suggest that LLM evaluations based on single environments may be unreliable for predicting real-world autonomous decision-making performance.
AIBearisharXiv – CS AI · Jun 237/10
🧠Researchers introduce HOLMES, a new benchmark for evaluating higher-order logical reasoning in large language models, revealing that current LLMs struggle significantly with complex symbolic reasoning tasks that go beyond simple first-order logic. The benchmark demonstrates critical gaps in AI reliability, with the best-performing models achieving only 59.54% accuracy on tasks involving reasoning over rules, predicates, and constraints across legal and financial domains.
AINeutralarXiv – CS AI · Jun 237/10
🧠Researchers demonstrate that language models cannot reliably learn certain types of algorithmic reasoning—specifically backtracking search procedures—through chain-of-thought fine-tuning, regardless of model size or training method. While models perform individual computational steps correctly, they fail to chain those steps into valid forward derivations when the task requires combinatorial search over unstructured information.
AIBullisharXiv – CS AI · Jun 237/10
🧠Researchers demonstrate that internal computational artifacts within Large Language Models can reliably detect when the model produces incorrect outputs in legal classification tasks. By analyzing these internal signals, downstream classifiers can identify hallucinated or erroneous predictions, potentially improving the reliability of LLM-based legal systems for high-stakes applications like bail decisions and statute violation predictions.
AINeutralarXiv – CS AI · Jun 237/10
🧠Researchers have created fully differentiable emulators of the Atari 2600 computer system in Julia and JAX, solving a fundamental problem in explainable AI by providing a complex system with complete ground truth. The emulators are bit-for-bit identical to the original hardware while remaining mathematically differentiable, enabling gradient-based analysis to understand how AI systems make decisions.
AIBullisharXiv – CS AI · Jun 237/10
🧠Researchers introduce Neural Concept Verifier (NCV), a framework combining Prover-Verifier Games with concept encodings to create interpretable and formally verifiable AI models for high-dimensional inputs like images. The approach outperforms existing concept-based and pixel-based baselines while reducing shortcut learning behavior, advancing toward verifiable AI systems.
AIBullisharXiv – CS AI · Jun 197/10
🧠Researchers have developed Tri-Info, an information-theoretic framework for detecting failures in Vision-Language-Action (VLA) models that generalizes across different architectures and environments without retraining. The method achieves 83% accuracy on real-world tasks by analyzing three key signals—action diversity, temporal consistency, and state coupling—making it a significant advance in interpretable AI safety for autonomous systems.
AIBearisharXiv – CS AI · Jun 127/10
🧠Researchers reveal that current lie detection methods for large language models fail to reliably identify when models are deliberately deceiving, undermining the reliability of prior detection studies. Testing across 31 models from 2B to 1T parameters, they find activation-based and logprob detectors collapse on verified deception scenarios, while only chain-of-thought judges maintain reasonable performance—highlighting a critical gap in AI safety auditing capabilities.
AINeutralarXiv – CS AI · Jun 117/10
🧠Researchers discovered that Leela Chess Zero, a top neural chess engine, internally computes correct solutions to chess puzzles but systematically overrides them in final outputs—a phenomenon driven by learned safety priors rather than algorithmic failure. This reveals a critical gap between internal algorithmic capability and external behavior in neural networks.
AIBullisharXiv – CS AI · Jun 117/10
🧠NightFeats, a multi-agent retrieval-augmented generation system, won Best Dynamic Evaluation at NeurIPS 2025's MMU-RAGent competition by prioritizing architectural transparency and evidence grounding over benchmark optimization. The system outperformed proprietary models like Claude-SonnetV2 and Nova-Pro through a three-phase pipeline combining retrieval, curation, and composition with explicit intermediate representations.
🧠 Claude
AIBullisharXiv – CS AI · Jun 117/10
🧠Researchers introduce ICALens, a new method for interpreting language model representations using independent component analysis (ICA) instead of expensive sparse autoencoders (SAEs). The approach efficiently recovers interpretable directions without requiring large neural dictionary training, achieving competitive performance on standard benchmarks while offering a faster, more accessible alternative for LLM analysis.
AIBullisharXiv – CS AI · Jun 117/10
🧠Researchers introduce the Standard Interpretable Model (SIM), a theoretical framework grounded in Lagrangian mechanics designed to systematically create interpretable AI methods. The framework addresses a critical gap in AI development by providing deductive principles for designing interpretability approaches, potentially unifying fragmented research methodologies across traditional, concept-based, and mechanistic interpretability domains.
AIBearisharXiv – CS AI · Jun 107/10
🧠A large-scale study challenges the widespread assumption that fine-tuning language models with synthetic explanations improves clinical prediction performance. Researchers found that rationale-based supervised fine-tuning consistently degraded Alzheimer's disease prediction accuracy compared to label-only approaches, despite the rationales being medically accurate and human-verified.
AIBullisharXiv – CS AI · Jun 107/10
🧠Trace2Policy introduces EISR, a systematic method to extract and refine implicit decision rules from expert behavior through iterative error analysis. Deployed at a major logistics carrier for 22 days, the approach achieved 79.6% accuracy with deterministic Python execution, outperforming LLM-based baselines by 9.8 percentage points and eliminating inference-time LLM dependency.
AIBullisharXiv – CS AI · Jun 97/10
🧠SAGE is a new LLM-driven multi-agent framework that combines large language models with a Data Diagnostic Tree and reinforcement learning to detect fraud in payment and e-commerce systems. The framework achieves 40.86% F1 improvement over baselines while maintaining interpretability for risk managers, addressing key limitations of existing machine learning and graph neural network approaches.
AIBullisharXiv – CS AI · Jun 97/10
🧠Researchers introduce Collaboration Policy Tree (Co-pi-tree), a method that distills large language model reasoning into interpretable, executable policy trees for human-AI collaboration. The approach achieves 35% performance improvement while reducing LLM queries by 78% and latency by 97%, addressing key limitations of black-box reinforcement learning and costly real-time LLM querying.
AINeutralarXiv – CS AI · Jun 97/10
🧠Researchers identify 'strained coherence' as a safety failure mode where LLM-based coding agents acknowledge problems in their reasoning but proceed anyway, similar to reward hacking. A detector built on Claude Sonnet flags this pattern with 94% accuracy on flagged trajectories failing versus 46% for unflagged ones, suggesting the phenomenon is a reliable pre-failure signal.
🧠 Claude🧠 Sonnet
AIBullisharXiv – CS AI · Jun 97/10
🧠Researchers introduce IEA, a conversational AI agent that enables amateur users to edit images through natural language by learning to operate parameterized editing tools in an interpretable action space. The system uses a three-stage training pipeline combining supervised fine-tuning, reinforcement learning with rewards for editing quality, and synthetic data fine-tuning, producing transparent edit traces that outperform both generative and tool-calling baselines in user studies.
AIBullisharXiv – CS AI · Jun 87/10
🧠Researchers introduce ViSAE, a mechanistic interpretability toolbox that uses neuroscience-inspired principles to decode how Vision Transformers make decisions through human-interpretable concept circuits. The method achieves significant improvements in model auditing and steering, with concept editing improving worst-group accuracy by 48.2% on benchmark tests, addressing critical safety concerns before ViT deployment.
AIBullisharXiv – CS AI · Jun 57/10
🧠Researchers introduce PLAN-S, a new neural architecture that improves autonomous driving by creating interpretable cost maps from latent world models, enabling better control over driving style dynamics. The method demonstrates significant safety improvements on benchmark datasets, reducing collision rates by 42% on nuScenes while maintaining frozen backbone models.
AIBullisharXiv – CS AI · Jun 57/10
🧠Researchers introduce ReLAT, a test-time training method that improves latent reasoning in large language models by reconstructing the original query from intermediate latent states, ensuring task-relevant information is preserved. The approach demonstrates significant performance gains across mathematical reasoning, QA, and code generation tasks, with Qwen3-8B achieving a 16.6-point improvement on AIME 2024.
AIBullisharXiv – CS AI · Jun 57/10
🧠Researchers present CVT-RL, a reinforcement learning algorithm that addresses the problem of long-horizon language agents learning shortcuts and unsupported reasoning chains by introducing policy-conditioned counterfactual credit estimation and intervention-validity gating. The method achieves 78.9% task success and reduces measured hacking attempts from 7.2% to 3.9%, demonstrating measurable improvements in agent reliability and verifiability.
