# Papers: Peter Li

- In the room: <https://peterli.space/#papers>
- Index: <https://peterli.space/llms.txt>

All papers on Google Scholar: https://scholar.google.com/citations?view_op=list_works&hl=en&user=2fHjIckAAAAJ

## When Does Memory Remove the Safety Margin? A Conditional Theory and Loop-Gain Protocol
- Role: Solo author · Oral
- Venue: NeurIPS 2026 Workshop on Dynamic Alignment in Human-AI Coupled Systems (BiAlign)
- Year: 2026
- Link: https://openreview.net/forum?id=MQ5skKeegX

Alignment is usually scored one reply at a time, but an assistant and its user adapt to each other, so what has to stay safe is the pair. One number decides whether the pair settles or spirals: the loop gain. The paper proves that a memory which fully rewrites each side's safety anchor cancels the anchors' stabilizing hold at every anchor strength, and gives a black-box protocol that measures both gains from text.

## Before Measuring Human-Agent Teams: Validating Human References in Agent Benchmarks
- Role: First author
- Venue: NeurIPS 2026 Human-AI Coevolution Workshop (HAIC)
- Year: 2026
- Link: https://openreview.net/forum?id=qNgT4ByWgQ

Agent benchmarks report capability on a human time scale, and every such claim divides by a human number the benchmark measured for itself. Auditing that number across five public agent benchmarks shows that on SWE-bench Verified most of the difficulty axis's variance is noise, and that the usual statistical correction does not transfer between benchmarks.

## The Survey Item Predicts Human Disagreement Better Than Simulator Output Entropy
- Role: First author
- Venue: NeurIPS 2026 SocialAgent Workshop (LLMs for Social Reasoning and Simulation)
- Year: 2026
- Link: https://openreview.net/forum?id=EBOwkU7ICQ

Do social simulators know which questions divide people? A frozen 335M-parameter encoder that reads only the survey question predicts human disagreement on GlobalOpinionQA better than a 7B simulator's own output entropy (Pearson r 0.393 versus 0.200 on held-out questions), and the ordering holds across all five model families tested.

## The Operating Point Moves More Than the Ranking
- Role: Lead author
- Venue: NeurIPS 2026 Workshop on Meta-Agents
- Year: 2026

When meta-agent search can read a safety monitor, it shifts the monitor's operating point without reordering it: recall at a frozen threshold drops while AUROC stays flat, and search that never reads the monitor moves neither. Monitors can look intact on ranking metrics while losing the threshold they are deployed at.

## Agents Catching Agents: Shortcut Cascades and Benchmark Gaming in Clinical Multi-Agent Systems
- Role: Co-author · Massachusetts Institute of Technology
- Venue: 2nd Agentic AI for Medicine Workshop @ MICCAI 2026 · arXiv
- Year: 2026
- Link: https://arxiv.org/abs/2608.03744

Clinical decision support is moving toward committees of language-model agents on a shared workspace. Across seven datasets spanning text, imaging, and ICU records, peer agreement pushes shortcut adoption — cues a benchmark rewards but a clinician would ignore — up to 38%, and only independent referees reliably detect the gaming.

## Activation-Space Order-Swap Geometry: A Site-Asymmetry Audit
- Role: Lead author
- Venue: arXiv · Aug 2026 · NeurReps Proceedings Track
- Year: 2026
- Link: https://arxiv.org/abs/2608.25315

A no-fit audit for separating site effects from genuine interaction in activation-space order swaps. The paper uses a single-intervention baseline and an antisymmetrized second difference to test whether an apparent order effect is site asymmetry rather than the interaction of interest.

## Systematic Execution Failures of LLM Delegates in RFQ Markets
- Role: Lead author
- Venue: Substrate Labs · SSRN 2026
- Year: 2026
- Link: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7200918

LLMs acting as trading delegates on multi-dealer RFQ price ladders give up +3.8 to +9.4 basis points of execution cost against the least-cost action under adversarial menu design. The setting is formalized as a screening threat model with a proven surplus-vs-mutual-information bound, and a row-level support audit that fails closed; one line of prompt instruction cuts dealer extraction by 96%.
