Zeyu (Steven) Zhang

PhD Student, Statistics and Data Science, Northwestern University

I am a fourth-year Ph.D. student in Statistics and Data Science at Northwestern University, working with Prof. Bradly C. Stadie. My main research interest is LLM forecasting.

My recent research asks how to make LLM backtests reliable. A central and hard obstacle is temporal leakage (look-ahead bias): knowledge from pretraining that lets a model “read the future” during a backtest. I also study the robustness of LLM forecasting systems: how they behave when the evidence they retrieve is manipulated. In addition, I am interested in the interpretation of LLM predictions—whether the prediction a model states reflects its internal reasoning and knowledge.

Research interests

LLM forecasting Trustworthy LLMs LLM interpretation

News

Nov 12, 2026	New preprint: One Fabrication Is Enough. A single fake news article reverses 74–82% of opposing LLM forecasts, even in the architectures of leading deployed forecasting systems.
Oct 29, 2026	Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted Scores is officially accepted at TMLR, with a decision of “accept as is.” That completes the backtesting trilogy alongside NeurIPS 2026 and the Findings of EMNLP 2026. This paper supplies the theory: leakage inflation cannot be identified from a passive backtest alone, and one external reference restores a leakage-adjusted score. Code and data are public.
Sep 24, 2026	TEMPO is accepted to NeurIPS 2026 (main track, 25.7% acceptance)! Rather than erasing what an LLM knows, TEMPO teaches it temporal discipline: a two-mode RL reward first drives post-cutoff leakage to zero, then optimizes prediction quality, with guarantees that leakage decreases monotonically along the way.
Aug 20, 2026	All Leaks Count, Some Count More is accepted to the Findings of EMNLP 2026 (14.3% acceptance)! The paper measures how much of an LLM’s decision-driving reasoning is temporally contaminated (Shapley-DCLR) and blocks it at inference time without retraining (TimeSPEC).
Aug 03, 2026	The capstone of our backtesting line is live on arXiv: Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted Scores proves the standard contamination check is uninformative and shows one defensible reference restores a leakage-adjusted score. In submission to TMLR, with code and data released.

Research

Temporal leakage in LLM backtesting

Backtesting evaluates a model’s prediction ability by asking historical questions as if it were forecasting from an earlier time. Its validity breaks when applied to LLMs: post-cutoff knowledge from pretraining leaks into the model’s reasoning and inflates the backtest score. My work treats this problem as a whole: a theory of when the inflation can be measured and corrected, an interpretable audit with an inference-time defense, and a training method that teaches the model temporal discipline.

Selected works in this line

Temporal leakage causal mechanism and the double-benefit law

Measure & correct the score Leakage Audit

Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted Scores

LLMs legitimately know more about times near their cutoff, so recency mimics leakage, and no passive backtest can separate recency and leakage from genuine skill. Leakage concentrates on hard questions, and even partial memorization captures most of the inflation (the “double benefit”). Two estimators—a known cutoff or a clean control model—yield a leakage-adjusted score, validated on twin models with planted leakage.

TMLR 2026

Measure & mitigate at inference TimeSPEC

All Leaks Count, Some Count More

Shapley-DCLR measures what fraction of a model’s decision-driving reasoning is contaminated. TimeSPEC interleaves temporally filtered retrieval with claim-level supervision so predictions rest entirely on pre-cutoff evidence, cutting leakage by 83–89% and reporting backtest performance honestly.

Findings of EMNLP 2026

Mitigate via training TEMPO

Temporal Enforcement via Mode-Separated Policy Optimization

GRPO post-training with a two-mode reward: a leakage mode drives post-cutoff claims to zero before a performance mode optimizes the task on clean evidence. Leakage falls from 2–13% to 0.6–3.7%, with performance up 6–13% where pre-cutoff signal is strong.

NeurIPS 2026

Attack and defense of LLM forecasting systems

Retrieval-based LLM forecasters search the news for evidence, and the wide search that makes them strong also makes them depend on the authenticity of every article they retrieve. We study the weakest attacker this design allows—one who knows only the forecasting question and publishes a single fabricated article—and show that even frontier models and leading deployed forecasting systems follow it, while accuracy hides the harm.

Selected works in this line

Directional attack: one fabricated article flips a stock forecast

Attack the evidence One Fabrication

One Fabrication Is Enough

An attacker who knows only the forecasting question publishes one fake news article. Once retrieved, it reverses 74–82% of opposing forecasts on every frontier model tested, and the architectures of two leading deployed forecasting systems retrieve it in about 90% of runs. A two-article version makes the model confidently wrong rather than uncertain, and the strongest published defenses are cheaply bypassed.

Preprint 2026

Interpretation: do explanations match decisions?

LLMs often give fluent explanations for their predictions, but fluency is not faithfulness. The hard question is whether the factors a model cites actually behave like they drive the decision—and how to check that without gradients or internal access. That makes interpretation both practically needed for trust and scientifically interesting as a window into model behavior.

Selected works in this line

Local decision surfaces LAMP

Extracting Local Decision Surfaces From Large Language Models

Fluent LLM explanations need not track what actually drives the prediction. LAMP treats the model’s own stated factors as coordinates and fits a local linear decision surface—without gradients or internal access—to test whether predictions move consistently when those factors change.

Spotlight paper in AISTATS 2026