Latent Cache Flow (LCF) jointly translates and compresses key-value pairs, reducing adapter size to 4% of Cache-to-Cache (C2C) while handling differing contexts. Experiments show a 13 MB LCF adapter outperforms a 956 MB C2C adapter on shared contexts and is 23% more accurate and 8.5x faster than text-based communication for different contexts.

Large reasoning language models (LRMs) produce long chain-of-thought trajectories containing reflection markers like 'wait', 'but', 'alternatively'. This paper reveals their distinct functional roles and timing of influence. PathCal is a training-free decoding controller that distinguishes marker types and intervenes only at locally uncertain states, achieving better efficiency-performance trade-off by reducing generation length while maintaining or improving accuracy.

This paper turns fundamental limits from Turing, Arrow, and No Free Lunch theorems into design rules, introducing the Deterministic Horizon: an accuracy ceiling set by architecture alone, beyond which no training can improve. Measured between 19 and 31 across 12 transformer architectures, fine-tuning on optimal-length traces recovers under 4%. The work extends to preference learning, retrieval pipelines, truthful auctions, and zero-knowledge verification, forming a catalogue of 16 specifications pairing computable boundaries, quantified violation costs, and constructive design rules.

EVE-Agent introduces evidence verifiability to self-evolving search agents by modifying the proposer–solver framework with an evidence verifier that rewards spans based on marginal accuracy gain. This ensures each training example includes a source-grounded span that explains why it should be trusted, leading to improved evidence-grounded correctness without human annotations.

This paper systematically develops the core of Mediative Fuzzy Logic and extends it to interval type-2, granular type-3, and quantum versions, establishing a convex aggregation operator controlled by hesitation and contradiction, proving soundness, paraconsistency, and conservativity, and demonstrating its transparent, conservative, and safety-first decision-making capabilities through an autonomous braking sensor fusion example.

ImProver 2 is a neurosymbolic framework for automated proof optimization in Lean 4. It combines a data-efficient expert-iteration pipeline with a scaffold that exposes formal structure alongside informal abstractions. A 7B-parameter model trained with this framework outperforms much larger models and competes with mid-tier frontier models, showing that small models can restructure research-level proofs effectively.

New research proposes A-LEMS, a framework that measures AI energy consumption per successful goal (EpG) rather than per inference, revealing that agentic workflows consume 4.33x more energy on average than linear baselines, with orchestration structure being the primary driver, but potentially more efficient for tool-augmented tasks.

SciAtlas integrates over 43 million papers from 26 disciplines into a knowledge graph with 157 million entities and 3 billion triplets, enabling AI agents to perform topologically-aware scientific reasoning and reduce logical hallucinations.

Research Math Agents (RMA) is an automated reasoning framework for research-level mathematical problems. It solves 8 out of 10 problems on the First Proof benchmark, outperforming GPT-5.2R and Aletheia through multi-agent collaboration and iterative refinement.

NeuroNL2LTL is a neurosymbolic framework that integrates learned translation with formal verification to convert natural language requirements into Linear Temporal Logic (LTL) specifications. By using a structure-preserving intermediate representation and verifier-in-the-loop training, it achieves 28% semantic equivalence and 86% satisfiability verification on over 200,000 requirements across multiple domains.

This paper introduces BOHM, a method to extract hierarchical attribution trees from routing weights in compound AI systems, requiring zero marginal cost and no access to component internals, providing multi-resolution attribution that correlates highly with SHAP but at a fraction of the cost.

This article clarifies often-confused AI agent terms like 'harness' (execution layer) and 'scaffold' (behavior-defining layer), explaining model, agent, tool use, sub-agents, and training concepts.