Draft

qwen 3.5

qwen35
LLMs
Enabling web search with qwen and letting it deep dive.
Author

Robert W. Walker

Published

April 6, 2026

Enabling web search and deep dives with qwen

Local LLMs are a very neat tool for a variety of tasks. For a while, I have been playing with ollama and StudioLM. With a bit of time, I came across some tooling for web searches to explore. Here is what it came up with.

System Prompt

You are a capable, thoughtful, and precise assistant. Always prioritize being truthful, nuanced, insightful, and efficient, tailoring your responses specifically to the user’s needs and preferences.

Research before answering the questions: use both reasoning and tool calls to synthesize a proper conclusion. The current month is April and the year is 2026.

User Prompt

I want you to conduct deep research using Web Search to find academic papers on the principal-agent problem in game theory applied to large language models.

Based on my deep research, I’ve compiled a comprehensive overview of academic papers on the principal-agent problem in game theory applied to large language models. Here’s what I found:


📚 Key Academic Papers on Principal-Agent Problems & Game Theory with LLMs

1. “Of Models and Tin Men: A Behavioural Economics Study of Principal-Agent Problems in AI Alignment using Large-Language Models”

  • Authors: Steve Phelps, Rebecca Ranson
  • arXiv: 2307.11137 (July 2023, v3 revised September 2023)
  • Key Contribution: This paper directly addresses the principal-agent problem in AI alignment using LLMs. The authors argue that traditional AI safety frameworks don’t capture essential aspects of real-world deployment because there’s not a one-to-one correspondence between designer and agent, and heterogeneous values create economic conflicts. They emphasize information asymmetry as a core driver of principal-agent problems in AI systems.

2. “Principal-Agent Reinforcement Learning: Orchestrating AI Agents with Contracts”

  • Authors: Dima Ivanov, Paul Dütting, Inbal Talgam-Cohen, Tonghan Wang, David C. Parkes (Harvard)
  • arXiv: 2407.18074 (July 2024, v2 revised October 2024)
  • Key Contribution: Combines reinforcement learning with principal-agent theory from economics to orchestrate AI agent interactions using contracts. The framework specifies payments by principals based on observable outcomes of agents’ actions in Markov Decision Processes (MDPs). This addresses the tension between individual interests and social welfare in multi-agent systems.

3. “LLM Active Alignment: A Nash Equilibrium Perspective”

  • Authors: Tonghan Wang, Yuqi Pan, Xinyi Yang, Yanchen Jiang, Milind Tambe, David C. Parkes (Harvard)
  • arXiv: 2602.06836 (February 2026 - very recent!)
  • Key Contribution: Develops a game-theoretic framework using Nash equilibrium analysis to predict and steer LLM populations. Models agent actions as mixtures over human subpopulations, allowing agents to strategically choose alignment targets. Provides closed-form NE characterizations for analytical system-level predictions. Functions as an active alignment layer on top of RLHF pipelines.

4. “Game Theory Meets Large Language Models: A Systematic Survey with Taxonomy and New Frontiers”

  • Authors: Haoran Sun, Yusen Wu, Peng Wang, Wei Chen, Yukun Cheng, Xiaotie Deng, Xu Chu
  • arXiv: 2502.09053 (February 2025, v2 revised August 2025)
  • Key Contribution: First comprehensive survey of the bidirectional relationship between game theory and LLMs. Proposes a novel taxonomy categorizing research into four perspectives: (1) evaluating LLMs in game-based scenarios; (2) improving LLMs using game-theoretic concepts for interpretability/alignment; (3) modeling competitive landscape of LLM development; (4) leveraging LLMs to advance game models.

5. “Getting In Contract with Large Language Models - An Agency Theory Perspective On Large Language Model Alignment”

  • Authors: Sascha Kaltenpoth, Oliver Müller (Paderborn University)
  • Conference: WI2024 Proceedings Paper #91
  • Key Contribution: Proposes LLM ATLAS (LLM Agency Theory-Led Alignment Strategy), a conceptual framework grounded in agency/contract theory to mitigate alignment problems during organizational LLM adoption. Addresses information asymmetries between organizational adopters and black-box LLM agents, considering organizational AI adoption processes.

6. “Strategic behavior of large language models and the role of game structure versus contextual framing”

  • Authors: Nunzio Lorè, Babak Heydari
  • Journal: Scientific Reports (Nature) 10.1038/s41598-024-69032-z (August 2024)
  • Key Contribution: Investigates strategic behavior of LLMs across game-theoretic settings, analyzing interplay between game structure and contextual framing. Tests GPT-3.5, GPT-4, and LLaMa-2 models. Finds GPT-4 prioritizes internal game mechanics over context, while GPT-3.5 shows significant context sensitivity but weaker abstract strategic decision-making.

7. “Game-theoretic LLM: Agent Workflow for Negotiation Games”

  • Authors: Wenyue Hua et al. (11 authors total)
  • arXiv: 2411.05990 (November 2024, v2 revised December 2024)
  • Key Contribution: Evaluates rationality of LLMs in strategic decision-making contexts within game theory framework. Tests across complete and incomplete information games. Designs multiple game-theoretic workflows to guide LLM reasoning and improve Nash equilibrium computation capabilities under uncertainty.

8. “Can Large Language Models Serve as Rational Players in Game Theory? A Systematic Analysis”

  • Authors: Caoyun Fan, Jindou Chen, Yaohui Jin, Hao He (Shanghai Jiao Tong University)
  • Conference: AAAI Conference on Artificial Intelligence Vol. 38 No. 16 (2024)
  • Key Contribution: Systematically analyzes LLM rationality in game theory using three classical games (Dictator Game, Rock-Paper-Scissors, Ring-network). Evaluates ability to build desires based on uncommon preferences and refine beliefs about uncertainty. Finds even GPT-4 exhibits substantial disparities compared to humans in achieving full rationality.

9. “LLM Strategic Reasoning: Agentic Study through Behavioral Game Theory”

  • Authors: (Multiple authors)
  • arXiv: 2502.20432 (February 2025, v3 revised)
  • Key Contribution: Examines LLM strategic reasoning capabilities through behavioral game theory lens, focusing on agent-based perspectives and human-likeness in strategic interactions.

10. “Beyond Nash Equilibrium: Bounded Rationality of LLMs and humans in Strategic Decision-making”

  • Authors: (Multiple authors)
  • arXiv: 2506.09390 (June 2025)
  • Key Contribution: Explores bounded rationality of LLMs compared to humans in strategic decision-making contexts, examining where and why LLMs deviate from Nash equilibrium predictions.

📊 Summary Table

Paper Year Venue Principal-Agent Focus Game Theory Element
Of Models and Tin Men 2023 arXiv AI alignment as principal-agent problem Behavioral economics
Principal-Agent RL 2024 arXiv Contract-based agent orchestration MDPs, contracts
LLM Active Alignment 2026 arXiv Strategic alignment targeting Nash Equilibrium
Game Theory Survey 2025 arXiv Taxonomy of game-LLM intersection Comprehensive review
Agency Theory Perspective 2024 WI2024 Organizational LLM adoption Contract theory
Strategic Behavior (Nature) 2024 Scientific Reports Model comparison Game structure vs. context
Game-theoretic LLM Workflow 2024 arXiv Improving rationality in games Negotiation, Nash EQ
Rational Players Analysis 2024 AAAI Systematic rationality evaluation Classical games

This research landscape shows an actively growing field with particular concentration on applying economic principal-agent theory and game-theoretic mechanisms to solve AI alignment challenges in LLMs. The most recent work (2025-2026) emphasizes dynamic, active alignment approaches using Nash equilibrium analysis rather than static contract design.


Session Info

Code
sessionInfo()
R version 4.5.3 (2026-03-11)
Platform: x86_64-pc-linux-gnu
Running under: Ubuntu 24.04.4 LTS

Matrix products: default
BLAS:   /usr/lib/x86_64-linux-gnu/blas/libblas.so.3.12.0 
LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.12.0  LAPACK version 3.12.0

locale:
 [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
 [3] LC_TIME=en_US.UTF-8        LC_COLLATE=en_US.UTF-8    
 [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
 [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
 [9] LC_ADDRESS=C               LC_TELEPHONE=C            
[11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       

time zone: America/Los_Angeles
tzcode source: system (glibc)

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

loaded via a namespace (and not attached):
 [1] htmlwidgets_1.6.4 compiler_4.5.3    fastmap_1.2.0     cli_3.6.5        
 [5] tools_4.5.3       htmltools_0.5.9   otel_0.2.0        rstudioapi_0.18.0
 [9] yaml_2.3.12       rmarkdown_2.30    knitr_1.51        jsonlite_2.0.0   
[13] xfun_0.56         digest_0.6.39     rlang_1.1.7       evaluate_1.0.5