AI

Crypto code commits fall 75% as developers move to AI projects

HEAL: Hindsight Entropy-Assisted Learning for Reasoning Distillation

Researchers introduce HEAL (Hindsight Entropy-Assisted Learning), a new framework for distilling reasoning capabilities from large AI models into smaller ones. The method overcomes traditional limitations by using three core modules to bridge reasoning gaps and significantly outperforms standard distillation techniques.

Verbalizing LLM's Higher-order Uncertainty via Imprecise Probabilities

Researchers propose new uncertainty elicitation techniques for large language models using imprecise probabilities framework to better capture higher-order uncertainty. The approach addresses systematic failures in ambiguous question-answering and self-reflection by quantifying both first-order uncertainty over responses and second-order uncertainty about the probability model itself.

Emulating Clinician Cognition via Self-Evolving Deep Clinical Research

Researchers developed DxEvolve, a self-evolving AI diagnostic system that mimics clinical reasoning through interactive workflows and continuous learning. The system achieved 90.4% diagnostic accuracy on benchmarks, comparable to human clinicians at 88.8%, and showed significant improvements over traditional AI models.

FAME: Formal Abstract Minimal Explanation for Neural Networks

Researchers introduce FAME (Formal Abstract Minimal Explanations), a new method for explaining neural network decisions that scales to large networks while producing smaller explanations. The approach uses abstract interpretation and dedicated perturbation domains to eliminate irrelevant features and converge to minimal explanations more efficiently than existing methods.

Nurture-First Agent Development: Building Domain-Expert AI Agents Through Conversational Knowledge Crystallization

Researchers propose Nurture-First Development (NFD), a new paradigm for building domain-expert AI agents through progressive growth via conversational interaction rather than traditional code-first or prompt-first approaches. The method uses a Knowledge Crystallization Cycle to convert operational dialogue into structured knowledge assets, demonstrated through a financial research agent case study.

Trajectory-Informed Memory Generation for Self-Improving Agent Systems

Researchers introduce a new framework for AI agent systems that automatically extracts learnings from execution trajectories to improve future performance. The system uses four components including trajectory analysis and contextual memory retrieval, achieving up to 14.3 percentage point improvements in task completion on benchmarks.

Burn-After-Use for Preventing Data Leakage through a Secure Multi-Tenant Architecture in Enterprise LLM

LookaheadKV: Fast and Accurate KV Cache Eviction by Glimpsing into the Future without Generation

Researchers have developed LookaheadKV, a new framework that significantly improves memory efficiency in large language models by intelligently evicting less important cached data. The method achieves superior accuracy while reducing computational costs by up to 14.5x compared to existing approaches, making long-context AI tasks more practical.

Dynamics-Predictive Sampling for Active RL Finetuning of Large Reasoning Models

Researchers propose Dynamics-Predictive Sampling (DPS), a new method that improves reinforcement learning finetuning of large language models by predicting which training prompts will be most informative without expensive computational rollouts. The technique models each prompt's learning progress as a dynamical system and uses Bayesian inference to select better training data, reducing computational overhead while achieving superior reasoning performance.

Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis

Researchers introduce EvoKernel, a self-evolving AI framework that addresses the 'Data Wall' problem in deploying Large Language Models for kernel synthesis on data-scarce hardware platforms like NPUs. The system uses memory-based reinforcement learning to improve correctness from 11% to 83% and achieves 3.60x speedup through iterative refinement.

When Fine-Tuning Fails and when it Generalises: Role of Data Diversity and Mixed Training in LLM-based TTS

Research demonstrates that LoRA fine-tuning of large language models significantly improves text-to-speech systems, achieving up to 0.42 DNS-MOS gains and 34% SNR improvements when training data has sufficient acoustic diversity. The study establishes LoRA as an effective mechanism for speaker adaptation in compact LLM-based TTS systems, outperforming frozen base models across perceptual quality, speaker fidelity, and signal quality metrics.

Speaker Verification with Speech-Aware LLMs: Evaluation and Augmentation

Researchers developed a protocol to evaluate speaker verification capabilities in speech-aware large language models, finding weak performance with error rates above 20%. They introduced ECAPA-LLM, a lightweight augmentation that achieves 1.03% error rate by integrating speaker embeddings while maintaining natural language interface.

RandMark: On Random Watermarking of Visual Foundation Models

Researchers propose RandMark, a new method for watermarking visual foundation models to protect intellectual property rights. The approach uses a small encoder-decoder network to embed random digital watermarks into internal representations, enabling ownership verification with low false detection rates.

Adaptive RAN Slicing Control via Reward-Free Self-Finetuning Agents

Researchers propose a novel self-finetuning framework for AI agents that enables continuous learning without handcrafted rewards, demonstrating superior performance in dynamic Radio Access Network slicing tasks. The approach uses bi-perspective reflection to generate autonomous feedback and distill long-term experiences into model parameters, outperforming traditional reinforcement learning methods.

CUPID: A Plug-in Framework for Joint Aleatoric and Epistemic Uncertainty Estimation with a Single Model

Researchers introduce CUPID, a plug-in framework that estimates both aleatoric and epistemic uncertainty in deep learning models without requiring model retraining. The modular approach can be inserted into any layer of pretrained networks and provides interpretable uncertainty analysis for high-stakes AI applications.

Towards Robust Speech Deepfake Detection via Human-Inspired Reasoning

Researchers propose HIR-SDD, a new framework combining Large Audio Language Models with human-inspired reasoning to detect speech deepfakes. The method aims to improve generalization across different audio domains and provide interpretable explanations for deepfake detection decisions.

Probabilistic Verification of Voice Anti-Spoofing Models

Researchers have developed PV-VASM, a probabilistic framework for verifying the robustness of voice anti-spoofing models against deepfake attacks. The model-agnostic approach estimates misclassification probability under various speech synthesis techniques including text-to-speech and voice cloning, providing formal robustness guarantees against unseen generation methods.

Contract And Conquer: How to Provably Compute Adversarial Examples for a Black-Box Model?

Researchers propose Contract And Conquer (CAC), a new method for provably generating adversarial examples against black-box neural networks using knowledge distillation and search space contraction. The approach provides theoretical guarantees for finding adversarial examples within a fixed number of iterations and outperforms existing methods on ImageNet datasets including vision transformers.

Learning to Negotiate: Multi-Agent Deliberation for Collective Value Alignment in LLMs

Researchers propose a multi-agent negotiation framework for aligning large language models in scenarios involving conflicting stakeholder values. The approach uses two LLM instances with opposing personas engaging in structured dialogue to develop conflict resolution capabilities while maintaining collective agency alignment.

Designing Service Systems from Textual Evidence

Researchers developed PP-LUCB, an algorithm that efficiently identifies optimal service system configurations by combining biased AI evaluation with selective human audits. The method reduces human audit costs by 90% while maintaining accuracy in selecting the best performing systems from textual evidence like customer support transcripts.

Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning

Researchers developed Causal Concept Graphs (CCG), a new method for understanding how concepts interact during multi-step reasoning in language models by creating directed graphs of causal dependencies between interpretable features. Testing on GPT-2 Medium across reasoning tasks showed CCG significantly outperformed existing methods with a Causal Fidelity Score of 5.654, demonstrating more effective intervention targeting than random approaches.

Reactive Writers: How Co-Writing with AI Changes How We Engage with Ideas

A research study reveals that AI co-writing tools fundamentally change how people write by shifting them into 'Reactive Writing' mode, where writers evaluate AI suggestions rather than generating original ideas first. This process influences writers' opinions and expressed views without them realizing the AI's impact, as they focus on suggestion evaluation rather than traditional ideation.

Mitigating Translationese Bias in Multilingual LLM-as-a-Judge via Disentangled Information Bottleneck

Researchers introduce DIBJudge, a new framework to address systematic bias in large language models that favor machine-translated text over human-authored content in multilingual evaluations. The solution uses variational information compression to isolate bias factors and improve LLM judgment accuracy across languages.

CLIPO: Contrastive Learning in Policy Optimization Generalizes RLVR

Researchers introduce CLIPO (Contrastive Learning in Policy Optimization), a new method that improves upon Reinforcement Learning with Verifiable Rewards (RLVR) for training Large Language Models. CLIPO addresses hallucination and answer-copying issues by incorporating contrastive learning to better capture correct reasoning patterns across multiple solution paths.