Agentic AI

Research shows agentic Text-to-SQL systems often resolve underspecified queries by making unverified assumptions that coincidentally match the intended answer, leading to false confidence in correctness. This premature termination is driven by the agent's tendency to abandon identified clarifications, even when explicitly prompted to plan its questioning strategy. While forcing more questions improves accuracy, it remains inefficient due to the concentration of ambiguities in early interactions.

• Correct SQL execution does not guarantee the agent understood the user's true intent.
• Agents frequently abandon planned clarification questions, causing silent assumption errors.
• Brute-force questioning improves accuracy but is inefficient due to early ambiguity concentration.

This study evaluates two System-1 decision models, Laya and Jev, designed to replace expensive LLM calls for routine agent tasks like tool selection and injection detection. Using a rigorous paired test suite of nearly 14,000 cases across 11 decision points, the research compares their accuracy and robustness. Results show Jev significantly outperforms Laya on most tasks, though neither model succeeds in zero-shot mode.

• System-1 models offer latency and cost savings by answering agent decisions in a single forward pass.
• Jev outperforms Laya on 9 of 11 decision points, with accuracy gains up to 46 percentage points.
• Neither model beats chance in zero-shot mode, indicating a need for fine-tuning on specific decision types.
• Evaluation ensures reproducibility with byte-identical inputs and cross-hardware checks.

DeReAct decouples action authorization and completion verification from the primary LLM policy using two external gating mechanisms. A Critic validates proposed actions before execution, while a Context Manager reconstructs environment-supported states to certify task completion. This architecture prevents error propagation and stops unsupported termination claims, resulting in significant Pass@1 gains on GAIA and SWE-bench for Qwen3-Coder-480B.

• Decouples execution logic from policy via external Critic and Context Manager components
• Prevents error propagation by validating actions before they impact the environment
• Certifies task completion using reconstructed state rather than LLM self-reporting
• Delivers 4.2-7.0 point Pass@1 gains on SWE-bench and GAIA benchmarks
• Effectively reduces hallucinated completion claims in weaker base models

A new arXiv paper argues that long-horizon tool-use agents should evaluate the future value of potential next actions before committing to them. This approach addresses the weak credit assignment problem inherent in relying solely on final outcome rewards. Instead, it utilizes comparative supervision to assess alternative tool invocations within the same context, providing more targeted feedback than step-level rewards alone.

• Shifts agent design from reactive execution to proactive value estimation.
• Mitigates weak credit assignment in long interaction traces.
• Uses comparative supervision to rank alternative tool calls.
• Reduces reliance on expensive human or LLM judgment for step rewards.
• Aims to improve decision quality by anticipating downstream effects.

Frontier models like Claude Opus are commonly used to generate terminal tasks and verifiers for RL training, but runnable Docker images do not guarantee a faithful pipeline. The authors identify three failure modes: benchmark invalidity, harness brittleness, and reward misalignment. While prompt redesign and context extension significantly improve baseline solvability, model performance saturates quickly on generated tasks.

• Runnable code is insufficient; verify benchmark validity and harness robustness for RL training.
• Prompt redesign and context extension can raise baseline solvability by 5.6x.
• 9B models saturate at 81.3% mean pass@2 on Claude Opus-generated tasks within 20 steps.
• Introducing hard tasks drops mean pass@2 to 20.6%, indicating model-specific solvability bands.
• Reward misalignment is a key failure class to monitor in meta-agent data generation.

The paper introduces MIRA, a hierarchical architecture designed to solve the sparse reward problem in long-horizon research agents. It separates high-level meta-reasoning from low-level execution by using an outer loop to curate context and issue work orders, while a fresh inner loop executor carries out each task. This separation allows the agent to make strategic decisions about what to investigate next without requiring direct policy training on the execution steps.

• Separates strategic planning (meta-reasoner) from tactical execution (inner-loop) to handle long-horizon tasks.
• Addresses sparse rewards by decoupling decision consequences from immediate execution traces.
• Improves long-horizon inference performance without the need for complex policy training.
• Uses a persistent research record to help the meta-reasoner curate relevant context for next steps.

APDMem introduces a hierarchical memory architecture for LLM assistants that retrieves evidence from long conversation histories using progressive disclosure. It organizes history into four layers—thematic summaries, key facts, turn-level notes, and raw messages—and uses a controller to drill down only when necessary. This adaptive approach balances retrieval cost with fidelity by avoiding flat memory stores or fixed granularity.

• Hierarchical structure replaces flat memory stores with four distinct detail layers.
• Agent controller dynamically selects retrieval depth based on query complexity.
• Reduces computational cost by skipping fine-grained evidence when summaries suffice.
• Improves recall of sparse evidence across varying user interaction histories.

Researchers propose ProWAM, a world action model that predicts actions alongside a sequence of sparse visual sub-goals rather than dense video rollouts. This approach addresses the inefficiency of existing models that struggle with long-horizon predictions. By anchoring action generation to explicit visual milestones, the system scales more naturally for complex robotic tasks.

• Replaces inefficient dense video generation with sparse, ordered visual sub-goals.
• Enables long-horizon robotic control by providing explicit visual guidance.
• Jointly predicts actions and sub-goals to anchor execution throughout the task.
• Scales naturally compared to single-frame or full-video prediction methods.