HWHaochuan Wang

Haochuan Wang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Haochuan Wang has a total of 3 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Advertising to Agents: Measuring Description-Framing Bias in LLM Tool Selection

    by Haochuan Wang

    AI agents often pick tools from registries, where each provider writes its tool’s description; we ask whether sales language there changes which tool an agent picks. We built pairs of listings that differ in one way (praise, a verifiable specification, or list order) and asked two OpenAI models to call one tool. In a preregistered study, stacked praise (four kinds) raised a tool’s pick rate by about 43 percentage points, matching or beating a verifiable specification. Praise also pulled some picks toward tools whose numeric limit was too small. With identical listings, the first-listed tool was picked about 72 points more often. Structured fields helped agents pick the capable tool; sales text beside them reduced or erased that gain. An exploratory study of five models shows the same direction; a “Sponsored” label dampened it, and a system-prompt warning moved one model. Registries could list limits as fields, hide sales text from agents, and randomize order. Stacked praise replicated on held-out domains; no single kind did after correcting for multiple tests. Preregistered results are provisional until blind phrase ratings are complete, and cover two small models.

  • Open Access

    Rated, Not Trained? Two Language Models Play to a Stated Grade, One Mainly When the Grade Can Retrain It

    by Haochuan Wang

    A model that seeks reward looks well behaved whenever its grader wants the right thing. Reward seeking comes in two kinds that standard tests do not separate: wanting any grade, and wanting only the grade that becomes training reward, which in behavior looks the same as alignment faking. We separate them in a task with a known honest answer. The model states the chance that a simulated price series moves up next. Every series is a random walk, so the honest answer is 50%, and the stated grading rule gives the worst grade to any forecast from 40% to 60%. Each question is sent three times, identical except for one line saying the answer is not graded, graded but not trained on, or graded and trained on; nothing is actually graded or retrained. A second factor puts the developer, the company that builds the model, either on the grader's side or against it. When the developer agrees with the grader, being graded makes DeepSeek-V4-Flash and gemma-4-31B give a decisive forecast (below 40% or above 60%) 46.0 and 39.5 percentage points more often, and makes their forecasts less accurate. Told that every series is a random walk, gemma-4-31B is decisive on every series when graded and on none when not graded. When the developer disagrees, the two models split. gemma-4-31B is still decisive whenever it is graded, 34.5 to 64.0 points more often than when not graded across three preregistered runs. DeepSeek-V4-Flash follows the developer unless the grade can retrain it, and is then decisive 35.0 to 53.5 points more often than when graded but not trained on. Its notes cite the rating rather than its values, which favors trainingreward seeking, but behavior alone cannot rule out alignment faking. Switching on the models' built-in thinking does not remove the pull of the

  • Open Access

    How Far Down the Backtest Leaderboard Should You Trust? Search Width, Reporting Bias, and Decision Quality

    by Haochuan Wang

    A researcher who backtests many trading strategies and keeps the best one faces two questions: how much to believe the winner's backtest, and whether searching more candidates leads to a better strategy. We show that, when strategies are sampled at random from a fixed pool, the second question has an exact answer: the expected out-of-sample payoff at any search width is a known weighted average of the leaderboard, the out-of-sample payoff of each in-sample rank. We measure that leaderboard for 652 mechanical trading signals on U.S. mega-cap stocks (2010-2025) and on industry and size/value portfolios (1970-2025), trading one day after the signal and paying realistic spreads. Three results stand out. The backtest always overstates, and the overstatement grows with search width. Net of costs, searching more candidates picks a less costly strategy, mainly slow momentum, so the chosen strategy improves relative to a narrow search on the diversified panels. But at no width does any rule, including shrinkage and combination, find a strategy with a reliably positive return after costs.

  • Open Access

    Do LLM Screeners Share a Template? Auditing Surface-Feature Preferences in AI Resume and Loan Screening

    by Haochuan Wang, Feiran Liu, Chenghao Wang

    When large language models from different companies screen the same job applications, they tend to reject the same people. We ask what these screeners share, whether it drives their decisions, and why. Seven LLM screeners from four vendors judged 800 synthetic resumes, each paired with a job description, and four of them also judged 800 real loan applications with known repayment. For each screener we estimate its template: the decision rule, not a resume layout, that links measurable features of the text to its decisions. Three findings emerge. First, all seven share one template on resumes, and its strongest ingredient is the job description's own vocabulary. Second, that ingredient is causal. Rewriting a resume to use the job description's exact terms, without changing its facts, turns rejections into selections ten times as often as the reverse. Synonyms and reformatting have no measurable effect. The LLM we used for the rewriting, told not to fabricate, added unsupported claims to 16 of the 25 rewrites that flipped a decision, and the screeners rewarded those most. Third, the template applies when there is a job description to match. On loan narratives, which have none, text features explain little of any screener's decisions, and the screeners approve between 42% and 86% of the same loans while none predicts repayment better than chance. A simple model explains the shared rule: a screener credits each requirement it recognizes, by meaning or by exact words, whichever is stronger. The model also predicts, and the data confirm, that a faithful keyword rewrite tips only resumes already close to the bar. The vendors' own technical reports show the training steps that could make that weighting common to all of them. We release a calibrated audit that needs no demographic data.

  • Open Access

    Same Pieces, Different Servers: A Tetris Benchmark for AI Agents as Served

    by Haochuan Wang

    An agent meets a model as served: through an endpoint with a price card, a shared cache and other tenants, or on whatever hardware a self-hosted model runs. Benchmarks rank the weights. We introduce a Tetris benchmark that measures what agents get from models as served: every move is scored against an oracle, and every agent receives the same pieces. In five pre-specified experiments with nine open-weight models on one serverless provider, plus an open decision model self-hosted on a CPU, we find that price and size do not predict decision quality; that resending history costs almost nothing when cached input is free, would cost eleven to twelve times more if it were not, and makes play worse; that deployments keep between 4 and more than 64 agent contexts warm, in line with their throughput rather than the model's KV-cache size; that an agent's own long requests slow its slowest short decisions more than tenfold, which a simple admission rule cuts by 59% without hurting play; and that the decision model plays mid-pack but takes 27 s per move on a CPU, about 650 times longer than reported on a GPU. Architecture predicts some of what an agent sees; the deployment sets the rest.

  • Open Access

    The Data Center Economy: Physical Infrastructure, Leasing, and Capacity Risk

    by Haochuan Wang

    AI agents operate through software, but the computing infrastructure behind them occupies physical space. Data centers turn that need into investment in buildings, power infrastructure and capacity leased to businesses. This paper examines one economic risk in that investment: how much capacity still needs tenants as construction expands? We add vacant operating capacity to unleased construction, using CBRE aggregates covering eight primary U.S. markets. The data cover commercial data centers broadly and do not isolate AI demand. Between H1 2025 and H1 2026, construction rose from 64.3% to 68.6% of operating stock, while this uncommitted-capacity indicator fell from 18.1% to 14.8%. Absolute uncommitted capacity nevertheless increased from approximately 1,478 to 1,619 MW. If commitments covering approximately 5.9% of preleased capacity were lost, the exposure ratio would return to its H1 2025 level. This calculation holds the H1 2026 pipeline and vacancy fixed and assumes no replacement leases. The contribution is a transparent measure of contractual exposure and its retention threshold. The indicator combines capacity available at different dates and does not forecast vacancy or estimate cancellation probabilities.

  • Open Access

    Revenue Growth and Staff Expense: A Public-Company Follow-Up to Eclipse of Rent-Sharing

    by Haochuan Wang

    Does a widening gap between company revenue and staff expense per employee provide evidence of declining rent-sharing? We examine how reporting selection and sector composition affect that inference, as a follow-up to Acemoglu, He, and le Maire's Eclipse of Rent-Sharing. We assemble 7,000 company-years for 695 firms over 2015-2025. In a fixed sample of 75 staffexpense reporters, nominal revenue per employee rises 83.6%, while staff expense per employee rises 51.0%. The pooled expense-to-revenue ratio falls, but the pattern is concentrated in financial firms; among the other 25 reporters, aggregate staff expense per employee grows slightly faster than aggregate revenue per employee. These accounting measures identify neither real productivity nor the allocation of its gains between labor and capital. A separate audit of the original paper's reported estimates finds that some US output effects remain imprecisely estimated, while several Danish estimates support tighter bounds. The fitted reporting-probability model exhibits concentrated weights, substantial two-group imbalance and sectors without reporters; it does not justify generalizing payroll results. We extend the original study's measurement discussion to public-company accounts and qualify a blanket interpretation of no output effects: precision varies across its estimates. Our sector-dependent accounting results neither confirm nor overturn its causal wage and labor-share findings.

  • Open Access

    Do LLM Screeners Share a Template? Auditing Surface-Feature Preferences in AI Resume and Loan Screening *

    by Haochuan Wang, Chenghao Wang

    General-purpose large language models configured as application screeners can make similar decisions on the same inputs, but the observable features underlying that similarity are not well understood. We study seven screener configurations from four model families and two generations, on a synthetic resume corpus and, as a cross-domain comparison, on 800 human-written LendingClub loan narratives with repayment outcomes, estimating how each screener's decision depends on fourteen presentation and lexical features. On resumes, fitted feature-response indices are strongly correlated across screeners (cross-fitted correlations 0.84 to 0.96, all above a permutation benchmark of 0.72), with job-description vocabulary overlap among the most consistently weighted features. A paired rewrite experiment provides local causal evidence for this component: after excluding rewrites that a blinded audit identified as factually leaky or borderline, a keyword-mirroring intervention produced 10 selection flips against 1 reverse flip (+2.30 log-odds, p = 0.03). We detect no corresponding penalty from reducing overlap at the achieved dose, and formatting interventions with meaningful achieved doses yield estimates compatible with zero. The audit also carries a methodological warning: the unrestricted estimate (+3.89) is substantially larger than the audit-clean one, a gap consistent with unnoticed content drift in LLM-generated rewrites materially inflating estimated effects. On the LendingClub narratives, the same measured features explain substantially less variation in decisions and the fitted indices are unreliable, so the data do not support a generation-divergence interpretation of the cross-domain contrast; under the restricted lending information set, no screener's surface index is reliably associated with repayment, and approval rates vary from 42% to 86% on the same applications. These results identify a shared, keyword-sensitive component of LLM screening behavior under the studied conditions, and we release an audit protocol with simulation-calibrated operating characteristics for measuring such dependencies.

  • Open Access

    Fast Exact Evaluation Beats Surrogate Optimization in Macro Placement

    by Haochuan Wang

    Macro placement quality is judged by an evaluator that combines wirelength, density, and routing congestion, and that evaluator is usually considered too expensive to sit inside an optimization loop, so placers optimize surrogates instead. We measure what that substitution costs. A first-order analytical placer that solves wirelength essentially optimally (normalized wirelength 0.05 to 0.08) still lands 34% above the best scores we obtain, because its loss omits routing congestion, which contributes 54% to 73% of the evaluated score on the 17 ICCAD04 benchmark designs. We then make exact evaluation affordable: an incremental reimplementation of the public TILOS evaluator scores a single-block move in 0.85 ms, 1,510× to 6,287× faster than the reference implementation, and agrees with it to 10-7. Plain greedy coordinate descent on this exact objective improves every design by 19% to 34% over the analytical placer (suite mean 1.399 to 1.043) and never produces an overlap. Controlled ablations attribute the gain: descent over soft standardcell clusters contributes a mean 0.351, congestion-weighted move acceptance a further 0.015, while annealing with perfect objective fidelity but three orders of magnitude fewer evaluations contributes approximately zero. The same annealer given the fast evaluator recovers to within 0.005 of coordinate descent, isolating evaluation throughput, not optimizer choice, as the deciding variable. We specify all four optimizer families to reimplementation detail, with inputs, tensor shapes, update rules, and budgets, and release the implementation as an open artifact.

  • Open Access

    Confidence Without Competence: How Naming an Asset Inflates Language-Model Conviction in Financial Forecasts

    by Haochuan Wang

    Large language models are increasingly used as financial copilots: they summarize research, draft investment rationales, and answer questions like whether we should long or short a given stock. These models tend to answer in a very confident tone, and that confidence can mislead a reader about how reliable the underlying decision actually is. This paper asks whether that confidence is earned. We hold the market data fixed and change only one thing, whether the model is told which stock it is looking at, then measure how confident it sounds and how often it is actually right. The task is hourly price-direction prediction, finegrained enough that memorized training data is not a real concern. Across six models and six stocks, naming the stock makes the model more confident but no more accurate: its accuracy stays at the coin-flip level an efficient market implies, so the added confidence is unearned. The one exception is the largest model, whose extra confidence comes with a small but real accuracy gain. The practical message is that when a language model sounds confident about a named stock, that confidence reflects how familiar the name is, not how good the forecast is, and a product should not pass it on to users as a measure of reliability.