2026Multi-objective Learning
SURF: Steering the Scalarization Weight to Uniformly Traverse the Pareto Front
NeurIPS 2026
Use the geometry of the Pareto front to choose weights that produce more evenly spaced model trade-offs.
Liuyuan Jiang · University of Rochester
The geometry of
learning and decisions.
Research directions
Competing objectives
Cover the range of useful trade-offs.
Equal steps in a preference weight can leave gaps on the Pareto front. I study how the front’s geometry can guide the choice of weights.
Model trade-offs
Same front. Different coverage.
Choosing an LLM summarizer. Each candidate balances summary quality and factual faithfulness. Each point is a different balance between these goals.
Higher scores are better. Click a point to explore a candidate.
Even weight steps can leave gaps between choices.
Summary quality
Faithfulness
Illustrative normalized scores. SURF studies quality and faithfulness rewards in LLM summarization. See the paper’s application.
Nested decisions
Optimize a decision through another problem.
In bilevel optimization, an outer decision changes an optimized inner response. I study how curvature, update rules, and moving constraints affect the cost of learning.
Price and response
The bakery sets a price. The customer responds.
BakerySets the price, x
CustomerChooses bread, y
The shop wants profit. The customer wants a purchase that is worth the price.
Add optional constraints
No extra constraints.
Customer buys, y*(x)
Bakery profit
Customer response y*(x)
How much bread the customer chooses at each price.
Preferred purchasePurchase with selected limits
Bakery profit
Profit(x, y) = (x − 1)y
The curve shows profit at each price, using the customer’s purchase y*(x).
Without extra limitsProfit with selected limitsSelected price
At $3 per kg, the customer buys 5 kg. The bakery earns $10.
The bakery chooses x. The customer determines y, so the bakery’s profit follows the response curve.
About this illustrative model
Without extra limits, the customer prefers 8 − x kg. This maximizes their modeled satisfaction minus the price paid. Bread costs the shop $1 per kg, so profit is (x − 1)y.
A budget limits spending: xy ≤ 12. Stock limits the amount bought: y ≤ 3.
Penalty methods · Rethink the cost of an updateSharper smoothness analysis, alternating updates, and flatness across constraint regimes.Smoothness · Follow the lower-level responseStudy reduced curvature to sharpen penalty-method iteration bounds.PBGD-Free · Remove the value-function loopUpper-level flatness bounds the bias from a simplified one-step-per-level update.BLOCC · Account for moving constraintsUse primal-dual information when the leader changes the follower’s feasible set.
Learning & behavior
Speech models and LLM agents.
My application work develops training targets for speech models and studies the timing and memory of LLM agents’ decisions.
Speech and agents
Learn from the model’s own features.
Enhanced labels evolve with the encoder; input-based anchors stabilize training.
Original BiRQ Figure 1. Paper source
Observe. Decide. Carry state forward.
The agent commits to long or flat before the next return is revealed.
Original RetailAgent Figure 2. Paper source
Publications & projects
2026Applications
RetailAgent: Structured Adverse Timing in Self-Conditioned Multimodal LLM Trading Agents
NeurIPS 2026 · IAB Workshop
A controlled audit of when LLM trading agents choose exposure, and what those decisions reveal.
2026Applications
BiRQ: Bi-Level Self-Labeling Random Quantization for Self-Supervised Speech Recognition
ICASSP 2026
Preprint first released in 2025.
Reuse a speech model’s own representations to improve its training targets, with raw-input anchors.
2025Bilevel Optimization
Efficient Penalty-Based Bilevel Methods: Improved Analysis, Novel Updates, and Flatness Condition
arXiv preprint · 2025
A broader analysis of alternating updates, flatness, and the cost of coupled constraints.
2025Bilevel Optimization
Improved Analysis of Penalty-Based Methods for Bilevel Optimization with Coupled Constraints
EUSIPCO 2025
Analyze curvature along the lower-level response to sharpen penalty-method iteration bounds.
2025Bilevel Optimization
Beyond Value Functions: Single-Loop Bilevel Optimization under Flatness Conditions
NeurIPS 2025
Use upper-level flatness to remove the value-function response computation from penalty-based bilevel training.
2024Bilevel Optimization
A Primal-Dual-Assisted Penalty Approach to Bilevel Optimization with Coupled Constraints
NeurIPS 2024
Optimize a leader’s decision when that decision also changes the follower’s feasible set.
About
Liuyuan Jiang
I am a PhD student in Electrical and Computer Engineering at the University of Rochester, advised by Prof. Lisha Chen. Previously, I was a PhD student at Rensselaer Polytechnic Institute, advised by Prof. Tianyi Chen.
I study how optimization shapes learning: navigating competing objectives, solving nested problems, and understanding model behavior.