AI Paper Digest · October 14, 2026
AI Paper Digest · October 14, 2026
This week’s digest covers three real papers: Kaiming He’s vision-native harness VISTA, the X-AgentRed red-teaming framework for personal agents, and RATED, a reliability-aware tool-retrieval method for agents.
1. VISTA: A Visual Harness for Reasoning in an Interactive World
- Title: VISTA: A Visual Harness for Reasoning in an Interactive World
- Authors / Institution: Qiushi Han, Keya Hu, Linlu Qiu, Cathy Wu, Kaiming He (MIT et al.)
- Problem: When multimodal models reason in interactive visual environments, the common approach is to abstract the environment into text or code (e.g., converting frames into text grids and having the model write programs to simulate rules). As environments grow more complex, object appearance, spatial relationships, and dynamics become increasingly hard to capture in text. Models also lack long-horizon visual memory: they cannot look back at, zoom into, or re-examine details from past observations.
- Method: VISTA, a vision-native harness. The model perceives the environment directly through visual observations and maintains a lossless visual memory that preserves all past frames in their original form. At reasoning time, the model actively retrieves prior observations and reorganizes its current visual input — turning “visual experience” into summonable context, with no retraining and no task-specific fine-tuning.
- Results: On ARC-AGI-3, pairing VISTA with Claude Opus 5.0 lifts Relative Human Action Efficiency (RHAE) from 40.68 to a perfect 100.00, completing all 25 public games using 57.4% fewer actions than first-time human participants; GPT-5.6 Sol also clears all games at 99. Across three additional benchmarks (browser games, mazes, matching puzzles), it substantially outperforms baselines that pair the same model with minimal harnesses, and it transfers to new visual environments with minimal adaptation.
- arXiv: https://arxiv.org/abs/2610.02200
2. X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems
- Title: X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems
- Authors / Institution: Yeonsung Jung, Zhiyuan Wang, Zifan Wang, Vítor Albiero, Benjamin Newman, Edoardo Debenedetti, Ivan Evtimov, Ninareh Mehrabi
- Problem: Personal agents increasingly connect to services across the web and apps through tools. Benign user requests can combine into harmful outcomes across services, but existing red-teaming methods mostly prompt LLMs to imagine harmful paths from task/environment descriptions. Generation is biased toward tool combinations seen during training, leaving novel cross-service interactions underexplored.
- Method: X-AgentRed, a self-evolving red-teaming framework that discovers harmful tool-call paths through a graph of tool dependencies. It analyzes each tool’s effects, potential harms, and execution preconditions, building a directed graph capturing how one tool’s effects satisfy another’s preconditions. Recipes sampled from the graph are instantiated as realistic scenarios agents execute; execution feedback drives self-improvement — diversifying harmful paths via tool substitution and refining preconditions at blocked steps so execution can continue.
- Results: Across four frontier agentic models, X-AgentRed achieves 1.77× the harmful-path rate of the STAC baseline and 2.99× the tool coverage on average, systematically surfacing compositional cross-service failures.
- arXiv: https://arxiv.org/abs/2610.13886
3. RATED: Reliability-Aware AI Agent Tool Retrieval via Execution Dynamics
- Title: RATED: Reliability-Aware AI Agent Tool Retrieval via Execution Dynamics
- Authors / Institution: Weiliang Chen, Song Liao
- Problem: AI agents increasingly depend on external tools, yet existing tool-retrieval methods optimize semantic relevance and ignore tool availability during execution. When APIs go down, time out, or lose permissions, the retriever keeps selecting unusable tools and tasks fail.
- Method: RATED combines a frozen dual-encoder with contrastive fine-tuning for semantic matching and a lightweight contextual-bandit module that re-ranks candidates based on real-time execution feedback — rapidly adapting to runtime tool failures without expensive model retraining. The authors also release EcoAgent, a benchmark of 5,172 real-world tools and over 210,000 query-tool pairs for offline retrieval and dynamic reliability evaluation.
- Results: RATED achieves a 52.2% relative MRR improvement over the ToolLLM API retriever on real-world queries, and cuts the fault-selection rate from 86.2% to 2.9% under a severe API-outage scenario with a 30% failure rate — all with negligible latency overhead (<1ms over encode-only inference). Code and dataset are open-sourced.
- arXiv: https://arxiv.org/abs/2610.13889
All articles on this blog are licensed under CC BY-NC-SA 4.0 unless otherwise stated.








