AI Agents

This paper addresses complex retrieval by shifting reasoning from query-time to index-time. The RL-Index framework uses an LLM to generate "rationales" for documents during indexing, encoding the latent reasoning needed to connect a query to the doc. The key is using Group Relative Policy Optimization (GRPO) with retrieval similarity as a reward signal, directly optimizing the index for downstream tasks. This improves both retrieval and QA performance while significantly cutting online inference latency.

This paper argues that bundling real-time search into a model is a poor architectural choice. The proposed Decoupled Search Grounding (DSG) pattern moves grounding outside the model into a vendor-agnostic gateway, exposing controls like provider routing, caching, and retrieval depth. On SimpleQA, DSG nearly matched native accuracy (86.1% vs 87.7%) with 91% lower search cost and 68% lower latency with a warm cache. It's a strong case for treating grounding as an optimizable interface.

A diagnostic study with a valuable negative result. The authors tried to improve agent policy by mining reusable skills from GUI interaction trajectories. While the mined skills were human-readable and correlated with ground-truth labels (0.95 purity on some clusters), they provided almost no benefit to a downstream policy. Training with mined skills only improved accuracy from 18.5% to 20.5% on one benchmark and had no effect on another. It's a sobering reality check for automated skill induction.

This paper argues for analyzing agent behavior by looking at execution trajectories, not just pass/fail rates. The authors formalize the "intent-execution gap" and use a new harness (Simple Strands Agent) to analyze 138k trajectories across popular benchmarks. They find that different frontier models exhibit distinct problem-solving styles—like edit frequency and testing strategies—that are hidden by aggregate pass@1 scores. This provides a finer-grained way to compare model behaviors for agentic tasks.

A production case study from DiDi on using an agentic pipeline to create user profiles for their ride-hailing dispatcher. ProfiLLM addresses platform scale: since most users are long-tail, an agent uses 27 analytical tools to mine platform data and generate profiles at the cluster level. Crucially, it uses a lightweight proxy for downstream utility to evaluate and refine candidate profiles via DPO, ensuring the generated text actually helps the production matching algorithm.

A framework for agentic vulnerability detection that forces the agent to make its reasoning explicit. When Code-Augur's agent deems code secure, it must commit the invariants behind that judgment as in-source assertions. A parallel fuzzer then tries to break these assertions. A triggered assertion reveals either a true vulnerability or a flawed assumption by the agent, creating a tight feedback loop that grounds the agent's analysis in verifiable specifications rather than opaque reasoning.

This paper presents DeXposure-Claw, an agentic system for supervising DeFi risk that deliberately constrains the LLM. Instead of letting the agent reason freely, it routes decisions through structured evidence. A time-series foundation model forecasts risk, deterministic monitors generate typed alerts from those forecasts, and only then does an LLM component create an auditable ticket. This "forecast-grounded" architecture is designed to prevent the false alarms and high-stakes errors endemic to free-form agent reasoning.

Proposes a new paradigm for alignment: instead of targeting abstract principles like "helpfulness," align models directly to structured, provider-authored specification documents. The SpecAlign framework operationalizes this by automatically synthesizing fine-grained preference pairs from the spec docs themselves. It uses rule annotation, controlled instantiation, and multi-agent adversarial synthesis to create training data capturing both compliant behavior and specific violations, enabling rapid adaptation to evolving policy requirements.

This work addresses the problem of web agent skill reuse. Instead of triggering skills based on instruction similarity or site metadata, SkillMigrator learns to transfer them by matching UI layout structure. It stores skills as "transferable interaction patterns" (TIPs)—the skill logic paired with a structural sketch of the UI when it was induced. At test time, it retrieves TIPs by layout similarity to the new page. This approach reduced the average LLM-action count by 8-10% on successful trajectories.

This paper proposes a middle path for creating multi-agent systems that bridges the gap between static prompting and costly fine-tuning. Skill-MAS avoids both by conceptualizing high-level orchestration as an evolvable "Meta-Skill." It uses a closed optimization loop that runs multiple trajectories under the current Meta-Skill, then uses selective reflection and contrastive analysis to distill experience into improved, generalizable strategic principles. The evolved Meta-Skills show strong transferability across tasks and base models.

A sobering benchmark for multi-turn agent safety in critical systems. NRT-Bench uses a simulated nuclear power plant control room run by a team of LLM agents. An adversary injects messages over multiple turns to induce failure. Adaptive attacks caused a critical safety function loss in 8.7% to 12.1% of sessions across four frontier models. Critically, vulnerabilities were model-specific: an attack that defeated one model often had no effect on another, and no single session defeated all four.

A concrete framework, C-Trace, for enforcing GDPR compliance at runtime for AI agents. It translates regulations like consent and purpose limitation into formal policy predicates over an agent's execution trace. A runtime monitor intercepts every tool call and model output, blocking non-compliant actions. Tested against attack dialogues designed to induce violations, the monitor kept the attack success rate under 12% and false positives under 16%, a significant improvement over baselines that rely on offline review.

Argues that natural language skill files (e.g., Markdown docs) are an inefficient way to guide agent behavior. SoftSkill proposes compressing a skill into a compact, trainable, continuous prefix instead. This "soft skill" is a small, learnable object (e.g., 32 virtual tokens) that initializes the model's state for a task. With a frozen Qwen3.5-4B, a SoftSkill prefix improved accuracy by 42.1 points on LiveMath over no-skill prompting, replacing thousands of Markdown tokens with a few latent ones.

This paper introduces JamSet and JamBench, a new dataset and benchmark for project-level code generation built on the Godot game engine. Sourced from over 240,000 Game Jam repositories, it provides 8,133 verified projects. The key finding from evaluating 9 frontier models is a "capability cliff" as projects scale: runtime pass rates plummeted from 80.4% on small projects to just 5.7% on large ones. Even when code agents improved compilation rates, runtime behavior didn't improve, pointing to architectural reasoning as the bottleneck.

A position paper arguing that current agent benchmarks based on aggregate-score leaderboards are misleading. The authors argue that rankings on these leaderboards do not transfer to out-of-distribution settings, citing recent competitions as direct evidence. They propose evaluating agent configurations based on "predictive validity"—the correlation between in-sample and out-of-sample rank—instead of just the in-sample score. This is a call to action for the field to build benchmarks that measure how well performance generalizes, not just how high it is.