Title: Cliff Tokens: Identifying Single-Token Failure Triggers in LLM Mathematical Reasoning

Source: http://arxiv.org/abs/2606.25524v1

Summary:

This paper identifies 'cliff tokens' as the exact single-token triggers that cause large language models to diverge into reasoning failures during multi-step mathematical tasks. By introducing a taxonomy of these failures and a targeted preference optimization method (Cliff-DPO), it establishes a foundational approach to diagnosing and improving LLM reasoning reliability.

Title: Agentic System as Compressor: Quantifying System Intelligence in Bits

Source: http://arxiv.org/abs/2606.25960v1

Summary:

This paper introduces a novel theoretical framework that quantifies agentic system intelligence through the lens of compression efficiency, linking agent capabilities like tool-use and search directly to codelength reduction. It provides a foundational methodology for analyzing, comparing, and optimizing multi-turn agentic workflows under budget and compute constraints.

Title: World Models in Pieces: Structural Certification for General Agents

Source: http://arxiv.org/abs/2606.24842v1

Summary:This paper introduces structural certification, a novel transition-local mathematical framework that guarantees goal-conditioned performance for general agents operating with segmented, non-universal world models. By proving tight error bounds, it establishes a foundational theory for localizing and certifying reliable long-horizon planning in complex environments.

Title: TRUSTMEM: Learning Trustworthy Memory Consolidation for LLM Agents with Long-Term Memory

Source: http://arxiv.org/abs/2606.25161v1

Summary:

This work is foundational as it presents a novel preference-guided reinforcement learning framework to optimize trustworthy memory consolidation for long-term LLM agents. By utilizing a Memory Transition Verifier to mitigate omission, corruption, and hallucination during updates, it introduces a robust architectural primitive for persistent agent state management.

Title: OpenThoughts-Agent: Data Recipes for Agentic Models

Source: http://arxiv.org/abs/2606.24855v1

Summary:

This paper establishes the first open, systematically ablated data curation pipeline and training recipes designed to generalize language models across diverse agentic benchmarks. By demonstrating strong training data scaling properties and releasing a high-performing 32B model, it provides a foundational blueprint for training next-generation agentic AI.

Title: Critique of Agent Model

Source: http://arxiv.org/abs/2606.23991v1

Summary:

This paper establishes a crucial conceptual boundary between workflow-scaffolded automation and true endogenous agency, introducing the novel Goal-Identity-Configurator (GIC) architecture. It provides a foundational blueprint for general-purpose agent models by integrating hierarchical goal decomposition, identity evolution, and simulative reasoning grounded in independent world models.

Title: Active Inference as the Test-Time Scaling Law for Physical AI Agents

Source: http://arxiv.org/abs/2606.22813v1

Summary:

This paper introduces a novel test-time scaling law for physical agents grounded in active inference and free energy minimization to handle out-of-distribution environments. By updating policies dynamically at test-time through variational inference, it unlocks continuous learning and scaling from real-world experience rather than just training data size.

Title: Self-Evolving Cognitive Framework via Causal World Modeling for Embodied Scientific Intelligence

Source: http://arxiv.org/abs/2606.22449v1

Summary:

This paper proposes a novel agentic cognitive framework that shifts embodied AI from simple predictive world modeling to self-evolving epistemic intelligence. It introduces integrated loops for causal discovery, intervention-driven reasoning, and continual cognitive refinement, which are essential for autonomous agent decision-making.

Title: All Routes Lead to Collapse

Source: http://arxiv.org/abs/2606.22325v1

Summary:

This paper presents a foundational geometric analysis demonstrating that representation collapse and attention sinks are inherent to content-based routing across diverse architectures, not just transformers. By showing this pathology persists in selective state-space models and recurrent mixers, it establishes crucial mathematical constraints for designing future architectural primitives.

Title: Learning the ARTS of Search for Automated Discovery

Source: http://arxiv.org/abs/2606.21891v1

Summary:

This paper introduces ARTS, a novel agentic reasoning framework that uses LLMs to navigate search spaces by diagnosing execution failures and selecting hypotheses. It also presents a reasoning breakthrough by using test-time training to distill search tree knowledge directly into model weights, allowing small open-source models to match closed frontier models at much lower cost.