Reasoning and planning
44 notes, newest first. Includes the vault tags reasoning, planning, long-horizon tasks, mathematical reasoning, chain-of-thought.
A Privileged Teacher That Only Scores Cannot Search: Critique-Conditioned On-Policy Self-Distillation for Reasoning
The study tests whether scoring signals in on-policy self-distillation localize errors and push probability away from erroneous student tokens. It reports that full-distribution KL and token log-ratio signals do…
GeoReform: Reflective Formalization Evolution for Multimodal Geometry Problem Solving
The preprint proposes GeoReform, which treats geometric formalization as an optimizable policy rather than a fixed parser output. It collects failed rollouts, diagnoses representation defects, and mutates how entities…
Cited but Not Consulted: A Counterfactual Audit of Legal Chain-of-Thought Faithfulness
The preprint tests whether legal models' named statutes or precedents causally support their verdicts. Holding case facts fixed, it substitutes the named authority and examines verdict changes and hidden-state dynamics…
ImproveAnyTask: An Autonomous Post-Training Harness for Iterative Model Self-Improvement
The preprint describes an autonomous post-training harness that attributes errors, selects update directions, and executes model updates under a limited compute budget. It combines metric- and case-level analysis…
Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis
The preprint introduces task-harness co-evolution for recursive reasoning-data synthesis. Online self-improvement turns solver failures into reusable skills, while post-task updates revise skills, prompts, and…
Cogentic: Multi-Agent Orchestration for Automated Proof Discovery
The preprint describes an orchestrator that assigns independent provers to competing proof directions, uses specialized adversarial verification components, and promotes confirmed intermediate results into a persistent…
Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning
The preprint introduces an inference-time harness in which a controller explicitly decides what partial work to reuse, whether to restart, when to stop, and which next work to dispatch. Workers perform task-level…
Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution
The preprint defines a Plan Declaration–Execution Gap and evaluates four declared planning modes: Predefined, Sequential, Hierarchical, and Search. Across four benchmarks and three LLMs, the abstract reports that only…
Shockingly Simple Self-retrospection Improves Agentic Models Without RL
The preprint introduces Retrospection-Only Fine-Tuning (ROFT), an online procedure in which an agent attempts a task, observes feedback, generates a retrospective explanation, and fine-tunes only on the explanation…
Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency
The preprint studies whether reasoning efficiency can improve without directly optimizing length or adding an inference-time stopping mechanism. It fine-tunes models to predict confidence at intermediate points in…
AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control
AD-WM is a joint-embedding world model designed for counterfactual model predictive control. It combines residual latent dynamics with action-recovery regularization so the model preserves action-dependent differences…
SAGE: Mitigating Long-Horizon Reasoning Biases via Topological Guidance
SAGE proposes Structural Admissibility-Guided Exploration for long-horizon reasoning under sparse rewards. It frames two failure modes—exploration bias toward locally plausible but unstable branches and compounding…
GRASP: Generating, Revising, and Assessing for Strategic Planning with Agentic AI
GRASP is a multi-stage planning framework with separate modules for global macro-guidelines, localized strategy exploration, and independent trajectory evaluation. The preprint reports gains over direct LLM planners on…
How Do Agent Harnesses Create Value? Planning Information and Release Control in Stateful LLM Agents
Across 265 matched cells in Retail experiments and an Airline pilot in τ²-bench, the paper compares task-specific plans with shuffled text, and evaluates a read-only terminal verifier. It reports a 7.17-point…
SkillAA: Attribution-Guided Skill-Graph Updating with Targeted Validation and Rollback
SkillAA represents skill applicability, execution, and composition in a unified graph. It contrasts successful and failed executions to route local repairs, then uses local and broader validation gates before…
STRETCH the Boundaries: A Unified Self-Taught Framework for Progressive LLM Evolution
STRETCH uses a dual-loop setup: a Scaffolder generates adaptive challenges near a model’s capability boundary, while a Learner improves solving trajectories with reinforcement learning. The abstract reports experiments…
Eureka: Task-Conditioned Meta-Agent Orchestration for Scientific Discovery
The paper presents Eureka, a Meta-Agent architecture that can complete long-horizon tasks by forming specialized Macro-Agents for specific tasks. The architecture allows for cost-benefit-gated evolution to update the…
ClawGym II: Exploring Black-Box RL on Agent Harness
Researchers present a new framework for reinforcement learning (RL) using complex agent harnesses. This framework is designed to optimize RL more stably and efficiently for long-horizon tasks. The researchers' method…
Neurosymbolic Embodied Agents
The authors propose a neurosymbolic agent that combines vision-language models with symbolic planning for executable long-horizon household tasks. The agent consists of two phases: the first involves a vision-language…
Better Slots, Better Worlds: Representation Quality & Robustness in Object-Centric World Models
The study investigates the performance of object-centric world models for visual model-predictive control, which enable agents to accomplish tasks through planning. The researchers find that these models perform better…
Better Slots, Better Worlds: Representation Quality & Robustness in Object-Centric World Models
The research focuses on object-centric world models (OCWMs) for artificial intelligence agents. The study compares OCWMs with traditional scene-centric models in terms of representation quality, planning success, and…
VAKRA: Evaluating Multi-Hop Reasoning Across APIs and Retrieval Under Tool-Use Policies
The researchers introduce VAKRA, a benchmark for evaluating the ability of agents to reason across structured APIs and document collections in enterprise settings. The benchmark includes over 8,000 executable APIs…
VibeLifeBench: Can Your Life Agent Be Proactive and Persistent in a Living World?
Researchers introduce a new benchmark called VibeLifeBench to evaluate large language model (LLM) agents' ability to be proactive and consistent in long-horizon tasks across ten everyday life domains. The benchmark…
Mismatch Matters: On-Policy Distillation Beyond Token Agreement
Researchers present TIDE, a new method to address a failure mode in on-policy distillation (OPD) called degenerate agreement. This issue arises when students exploit repetitive loops to achieve high token agreement…
Privileged, but Biased: How PI-Conditioned Teachers Break Self-Distillation
The research investigates self-distillation (SD), a method used in reinforcement learning with verifiable rewards, where a teacher conditioned on privileged information supplies supervision to a student. The study…
The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation
The paper focuses on improving the multi-turn long-horizon planning for foundation model agents. The authors introduce a controlled environment to study the acquisition, shaping, and integration of planning abilities…
Mask-Aware Policy Gradients for Diffusion Language Models
The paper presents a method for improving the performance of large language models in tasks like mathematical reasoning and coding by using reinforcement learning. This method addresses the challenge of extending…
OmniaBench: Benchmarking General AI Agents Across Diverse Scenarios
A research paper titled 'OmniaBench' introduces a new benchmark for evaluating the capabilities of general AI agents across various scenarios, using diverse tasks drawn from app stores, product documents, industry…
AIMO Interpretability Challenge
The AIMO Interpretability Challenge is a competition proposed by researchers aimed at distinguishing robust from spurious reasoning in advanced mathematical language models. The challenge focuses on evaluating the…
StructAgent: Harness Long-horizon Digital Agents with Unified Causal Structure
In this paper, researchers introduce StructAgent, a framework designed to improve long-horizon digital agents' performance in tasks with evolving contexts and complex workflows. The framework introduces a unified state…
OpenCoF: Learning to Reason Through Video Generation
The paper introduces OpenCoF, a framework that includes a new dataset and a fine-tuned video model for improving Chain-of-Frame (CoF) reasoning in video generation. The authors argue that existing models lack specific…
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
Researchers introduce a proactive memory agent for long-horizon tasks, aiming to improve decision-making in scenarios where relevant state information can become buried over time. The memory agent updates a structured…
Reasoning Consistency Scanning: A Framework for Auditing Chain-of-Thought Validity in AI Safety Evaluations
The paper introduces a method called Reasoning Consistency Scanning to assess the logical consistency between stated reasoning and answers in AI safety evaluation transcripts. The method can be applied to various AI…
TREK: Distill to Explore, Reinforce to Refine
TREK is a procedure that uses distillation for exploration support expansion to improve policy optimization on tasks where the current policy stalls on hard prompts. It identifies low-performing prompts, generates…
AutoMem: Automated Learning of Memory as a Cognitive Skill
Researchers developed AutoMem, a framework for improving a language model's memory management skills in long-horizon games. By treating memory as a trainable skill, they allowed the model to decide how to manage its…
AdaJEPA: An Adaptive Latent World Model
The paper proposes a new adaptive latent world model called AdaJEPA that can improve planning accuracy by continuously recalibrating the world model during test-time, particularly when dealing with distribution shift…
Freeform Preference Learning for Robotic Manipulation
Researchers introduce Freeform Preference Learning (FPL), a method for training robotic policies based on natural-language preferences provided by humans, rather than binary labels. The method uses the preferences to…
QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents
Researchers introduce QVal, a tool that evaluates dense supervision signals for long-horizon Learning with Long Memory (LLM) agents without the need for training. This allows for a fair comparison between different…
QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents
A research paper introduces QVal, a tool for evaluating dense supervision signals in long-horizon tasks for LLM agents without the need for training. The tool measures how well a method's score aligns with Q-values of…
Advancing Omnimodal Embodied Agents from Isolated Skills to Everyday Physical Autonomy
Researchers have developed a new framework called OmniAct for building embodied agents that can operate in unstructured environments. This framework aims to address the challenges of managing both cyber and physical…
Advancing Omnimodal Embodied Agents from Isolated Skills to Everyday Physical Autonomy
The paper presents OmniAct, a framework designed to create persistent embodied agents that can operate autonomously in unstructured environments. It addresses the challenges of integrating cyber (APIs, IoT) and…
Advancing Omnimodal Embodied Agents from Isolated Skills to Everyday Physical Autonomy
Researchers introduce OmniAct, a framework designed for creating persistent embodied agents in unstructured environments. The framework is built on a hierarchical asynchronous architecture that separates planning…
Empowering GUI Agents via Autonomous Experience Exploration and Hindsight Experience Utilization for Task Planning
The research introduces a method called Planning Experience Exploration and Utilization (PEEU) to improve the task planning capabilities of smaller open-source Multi-Language Model Learning Machines (MLLMs). These…
Empowering GUI Agents via Autonomous Experience Exploration and Hindsight Experience Utilization for Task Planning
Researchers introduce a new method called PEEU (Planning Experience Exploration and Utilization) to improve the task planning capabilities of small, open-source Multimodal Language Learning Models (MLLMs). These models…
