nithilan karthik nithilan@managent.tech UPDATES: new research

02

research

Four projects. The first is the research behind managent, accepted to the NeurIPS 2026 VeriCodeGen workshop. Two are at MIT LIDS. The last one is independent. The PDFs on this page are the accepted papers.

01

managent · accepted, NeurIPS 2026 VeriCodeGen

Asymmetric Repository Lineage Modeling and Verifier-Guided Coordination in Concurrent AI Coding Agents

With Arjun Subramanian and George Xu. Accepted to the NeurIPS 2026 workshop on AI for Verifiable Coding, December 12 in Atlanta. This is the research managent comes out of.

The earlier measurement is on arXiv: AI Agent Pull Requests on GitHub: Frequency, Structure, and Merge Conflict Rates. This paper is about what you can know before you check. A file list is a weak signal. On 715 replayed pairs, 79 of 167 textual conflicts, 47.3%, happen even though the two pull requests edited different files. The three-way merge looks at branch history, not just the files each PR lists. A lineage rule ranks those pairs well. The checker is still git merge-tree. When that replay is cheap, run it on everything. A score only decides what to check next.

02

mit lids · ongoing

cloning

At MIT LIDS I’m working with a professor on fair group selection when the participants aren’t people. They’re synthetic agents: LLM personas and civic simulations you can copy cheaply.

Most fairness rules assume an identity is scarce. Flood a category with near-copies of the same agent and “equal chance” stops meaning much. We’re looking for mechanisms that still hit a representation goal when the population can clone itself. Aiming at AAMAS 2027, so I’m staying light on the method.

03

mit lids · accepted, ACM ICAIF workshop

Verify Before You Sign: Benchmark Sensitivity and Tool-Output Deference in LLM Portfolio Advice

With Arjun Subramanian. A model can hand you a portfolio that looks reasonable and still be unsafe to sign. We audited two separate failures on 16 synthetic households: how far the score moves when you change the benchmark, and whether the model just copies a printed portfolio even when it breaks a hard cash floor.

gpt-5.6-luna copied 14 of 16 planted portfolios that violated the floor. claude-sonnet-5.5 copied none of them. The rule we argue for is strict: report the benchmark sensitivity, and enforce hard constraints outside the model before anyone signs.

04

independent · accepted, ACM WoAIS ’26

Static Before Smart: Exact Cost Accounting and Baselines for Pay-per-Token LLM Routing

With Arjun Subramanian. Accepted to the Second International Workshop on AI and Serverless, December 14, 2026, in Tarragona. People often score an LLM router by multiplying how often it picks each model by that model’s average cost. That shortcut breaks once the choice depends on the query, because long requests and expensive models travel together.

We write down the exact request-level cost, then compare a router to a static mix that spends the same money. On the held-out test, the query-only routers didn’t beat that static baseline. A hindsight oracle still had room left. Before calling a router a win, match it on real dollars.