Aarush Sinha
Publications
Evaluation Cards: An Interpretive Layer for AI Evaluation Reporting
AI evaluation results are produced at scale but reported inconsistently across leaderboards, model cards, benchmark papers, and company blogs. The cost is interpretive: readers cannot reliably compare results across sources, identify what a report omits, or trace an aggregate claim to its underlying evidence. Recent efforts address isolated components but leave three gaps: they cover only narrow slices of the evaluation lifecycle and do not compose into a single interpretable record; they specify static representations that do not differentiate the questions different stakeholders bring to the same evidence; and they remain proposals on paper, lacking the extraction infrastructure required for adoption at scale. We present \EvalCards{}, an operational reporting layer that composes benchmark metadata, evaluation run data, and model metadata into a unified record. We (1) derive a reporting schema from a structured review of 52 papers and 10 stakeholder interviews, (2) implement four interpretive signals (reproducibility, documentation completeness, provenance and risk, and score comparability), rendered through reader modes calibrated to research and non-research audiences, and (3) deploy a monitoring tool that applies \EvalCards{} across 5,816 models, 635 benchmarks, and 101,843 results, surfacing systematic gaps in current reporting practice.
CONSCIENTIA: Can LLM Agents Learn to Strategize? Emergent Deception and Trust in a Multi-Agent NYC Simulation
As large language models (LLMs) are increasingly deployed as autonomous agents, understanding how strategic behavior emerges in multi-agent environments has become an important alignment challenge. We take a neutral empirical stance and construct a controlled environment in which strategic behavior can be directly observed and measured. We introduce a large-scale multi-agent simulation in a simplified model of New York City, where LLM-driven agents interact under opposing incentives. Blue agents aim to reach their destinations efficiently, while Red agents attempt to divert them toward billboard-heavy routes using persuasive language to maximize advertising revenue. Hidden identities make navigation socially mediated, forcing agents to decide when to trust or deceive. We study policy learning through an iterative simulation pipeline that updates agent policies across repeated interaction rounds using Kahneman-Tversky Optimization (KTO). Blue agents are optimized to reduce billboard exposure while preserving navigation efficiency, whereas Red agents adapt to exploit remaining weaknesses. Across iterations, the best Blue policy improves task success from 46.0% to 57.3%, although susceptibility remains high at 70.7%. Later policies exhibit stronger selective cooperation while preserving trajectory efficiency. However, a persistent safety-helpfulness trade-off remains: policies that better resist adversarial steering do not simultaneously maximize task completion. Overall, our results show that LLM agents can exhibit limited strategic behavior, including selective trust and deception, while remaining highly vulnerable to adversarial persuasion.
$\mathrm{ECI}_{\mathrm{sem}}$: Semantic Residual Effective Contrastive Information for Evaluating Hard Negatives
Hard-negative source selection for dense retrieval is usually decided only after fine-tuning and downstream evaluation. We propose ECIsem, a validity-weighted diagnostic that ranks candidate hard-negative sources using frozen target-encoder embeddings. ECIsem is training-free, not label-free: each scored example requires a query, a labeled positive, and an explicit candidate negative. Each negative is weighted by target consistency, semantic locality, and lexical residuality, and the source score aggregates weighted residual directions through a log-determinant. At the scale we study, the log-determinant operates in its linear regime, so the score reduces to the mean validity weight; we report this reduction as a finding, and gate ablations show that a single lexical scalar, mean lexical residuality, reproduces the transfer ordering in every block. On MS MARCO negative sources, in-family ECIsem recovers the aggregate BEIR transfer ordering across DistilBERT, E5-base, and Contriever. Constructed controls expose an identifiability limit: a plausible false negative is non-inverted, near the positive, and lexically novel, exactly the profile the gates reward, and across twelve weightings of the three gates and three backbones only one of thirty-six cells is safe, and it is not transfer aligned, a limit that extends to margin-style filters built from the same statistics. The same gate statistics separate the controls sharply when used as audit features rather than as rewards, so we pair the score with a reported audit profile instead of a reweighting.