Bob Yan

Reading

Readings, by the question.

The stack behind evaluation for intelligence that is not confined to human language. Each section is one question, then the papers that take it up. Safety here is more than one measure: what the model is doing on the inside, and whether a complicated task or scenario still holds.

Vocabulary and control

A model can hold a concept that ordinary words do not name. These papers treat that gap as a control problem: a new term, a learned neologism, an instruction the model was not tuned on, and a trace that may not say what the reward paid for.

8 papers

  1. Hewitt, Geirhos, and Kim Position: We Can’t Understand AI Using our Existing Vocabulary 2025
  2. Hewitt, Tafjord, Geirhos, and Kim Neologism Learning for Controllability and Self-Verbalization 2026
  3. Kim, Hewitt, Nanda, Fiedel, and Tafjord Because we have LLMs, we Can and Should Pursue Agentic Interpretability 2025
  4. Hewitt, Liu, Liang, and Manning Instruction Following without Instruction Tuning 2024
  5. Ma and Hewitt Improving Parametric Knowledge Access in Reasoning Language Models 2026
  6. Morgulis and Hewitt Subliminal Steering: Stronger Encoding of Hidden Signals 2026
  7. Tianyi Lorena Yan and John Hewitt Reward Misspecification Degrades Chain-of-Thought Monitorability 2026
  8. Wijaya et al. Learning Translations via Matrix Completion 2017

Provenance, persona, and roles

Social behavior has a source, and a role can be a direction in the model rather than a line in the prompt. The question is whether that direction stays local, or leaks into the next client and the next decision.

5 papers

  1. Matlin et al. Capability Provenance in Language Models: A Case Study in Social Reasoning 2026
  2. Song et al. Role Steering of Language Models for Social Simulations 2026
  3. Gupta, Krier, and Matlin When Goals Leak Across Relationships: How Cached Goal Representations Shape Cross-Client Decisions 2026
  4. Okamoto, Erol, and Matlin Trust by Design: Skill Profiles for Transparent, Cost-Aware LLM Routing 2026
  5. Matlin et al. Do Language Models Agree with Human Perceptions of Suspense in Stories? 2025

Mechanistic interpretability and misalignment

Persona, emotion, and reward leave traces inside the model. These papers look for the direction, the feature, or the training step that carries the behavior, including when a narrow fine-tune changes the model well outside the task.

5 papers

  1. Chen et al. Persona Vectors: Monitoring and Controlling Character Traits in Language Models 2025
  2. Betley et al. Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs 2025
  3. Deng et al. Directional Alignment Mitigates Reward Hacking in Reinforcement Learning for Language Models 2026
  4. Shu, Singh, and ElSherief Why Are Some Emotions Harder for LLMs? Uncovering the Causal Mechanisms of Emotion Inference via Sparse Autoencoders 2026
  5. Quirke et al. Bergson: An Open Source Library for Data Attribution 2026

Finance

A plausible sentence about markets is not a result. These papers ask whether a model can do the financial task it was given, and what happens once it can place an order, call a tool, or act in a market.

7 papers

  1. Matlin et al. Finance Language Model Evaluation (FLaME) 2025
  2. Matlin et al. Financial Instruction Following Evaluation (FIFE) 2025
  3. Matlin et al. FinForge: Semi-Synthetic Financial Benchmark Generation 2026
  4. Chen et al. FinQA: A Dataset of Numerical Reasoning over Financial Data 2021
  5. Yang et al. FinVault: Benchmarking Financial Agent Safety in Execution-Grounded Environments 2026
  6. Ding et al. Beyond Knowledge to Agency: Evaluating Expertise, Autonomy, and Integrity in Finance with CNFinBench 2025
  7. Lei et al. Strategic Exploitation in LLM Agent Markets: A Simulation Framework for E-Commerce Trust 2026

Neologisms, languages, and tasks

Language and the world both move. These papers test new words, a changed timeline, stance across languages, and tasks whose difficulty is the order of the steps or the coordination of many agents.

9 papers

  1. Zheng, Ritter, and Xu NEO-BENCH: Evaluating Robustness of Large Language Models with Neologisms 2024
  2. Zheng et al. Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs 2026
  3. Zheng, Das, Ritter, and Xu Probabilistic Reasoning with LLMs for k-anonymity Estimation 2025
  4. Lavrouk et al. Stanceosaurus 2.0: Classifying Stance Towards Russian and Spanish Misinformation 2024
  5. Zheng, Baheti, Naous, Xu, and Ritter Stanceosaurus: Classifying Stance Towards Multicultural Misinformation 2022
  6. Lee et al. KoNeoBench: A Curated Evaluation Dataset for LLM Understanding of Korean Neologisms 2026
  7. Naous et al. Camellia: Benchmarking Cultural Biases in LLMs for Asian Languages 2025
  8. Gao et al. SwarmBench: Can Large Language Models Act as Agent Swarm Orchestrators? 2026
  9. Le, Guo, Xu, and Ritter Improved Instruction Ordering in Recipe-Grounded Conversation 2023