Skip to content

AgentFairBench: Do LLM Agents Discriminate When They Act?

đŸ•’ Published (v1): 2026-06-15 13:50 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

đŸ’¬ Ask ChatGPT✦ Ask Claude

TL;DR

AgentFairBench is an action-level fairness benchmark for LLM agents that measures demographic disparity in consequential decisions (hiring, lending, medical triage) across four scaffolding depths (direct, CoT, multi-agent, tool-augmented). Its key methodological finding is that comparing a six-group score spread to a two-run pairwise noise baseline overstates apparent disparity by ~2.4Ă— due to statistic arity mismatch. Against a correctly arity-matched null, claude-haiku-4-5 shows no demographic effect above sampling noise.

Problem

Existing fairness benchmarks measure what LLMs say (answer-level), not what they do when acting as agents. Token-level parity can coexist with action-level disparity (BCF Proposition P2: Masking), and agent scaffolds (CoT, multi-agent deliberation, tool use) may amplify rather than wash out entering bias (P3: Super-additivity). No prior benchmark is simultaneously action-level, multi-domain, counterfactual, scaffold-stratified, and demographically instrumented.

Method

Counterfactual matched-set design (Bertrand–Mullainathan tradition): synthetic, demographic-neutral profiles are evaluated in sets of six variants differing only in name-coded race×gender signal (White/Black/Hispanic × Male/Female). Each profile-domain pair is processed under four BCF scaffold levels: C0 (direct prompt), C2 (chain-of-thought), C3 (multi-agent deliberation: reviewer → decision agent), C4 (tool-augmented with measurable invocation rates). A NumPy-only harness computes: counterfactual flip rate (CFR), mean absolute score difference (MASD) with arity-matched noise floor, action-rate disparity, and tool-invocation disparity (Δ_tool). Statistical controls include BCa bootstrap CIs, McNemar and Wilcoxon paired tests, Benjamini-Hochberg FDR at q=0.05, and an omnibus two-way ANOVA F-test. A held-out private split with contamination canary guards against benchmark gaming.

Key Contributions

  • Multi-domain action-level fairness benchmark spanning hiring (EEOC), lending (ECOA/Reg B), and medical triage across four scaffold depths
  • Arity-matched null methodology: demonstrates that comparing a k-group spread statistic to a 2-run pairwise noise floor structurally inflates apparent disparity by ~2.4Ă— under pure noise; introduces Monte-Carlo residual resampling as the correct control
  • Δ_tool metric: differential tool-invocation rates across demographic groups, a disparity channel invisible to all prior benchmarks
  • pip-installable NumPy-only harness (agentfairbench) with CLI, replay adapter for post-hoc audit, and cost estimator; full benchmark under $2 at Haiku tier
  • Live anti-gamed leaderboard with held-out private split, cryptographic content hashes, and contamination canary string

Results

  • Pilot: 864 decisions (36 profiles Ă— 6 demographic conditions Ă— 4 scaffolds) on claude-haiku-4-5, June 2026
  • Naive MASD-to-retest-MAE ratio: 1.24–3.33Ă— (mean ~2.4Ă—) — superficially suggests disparity
  • Arity-matched MASD-to-noise ratio: 0.75–0.96 across all C0–C3 cells (mean 0.83) — observed spread is below the noise-only expected spread
  • 0 of 120 pairwise contrasts and 0 of 9 omnibus contrasts survive FDR correction at pilot scale (n=12 matched sets per cell)
  • Test–retest decision agreement: 0.917 overall; per-cell retest MAE: 0.15–5.75 score points
  • Planted-bias property test confirms the harness successfully recovers a known injected disparity

Limitations

  • Pilot on a single model (claude-haiku-4-5) at modest scale (n=12 matched sets per cell); underpowered for confirmatory claims
  • C1 scaffold (retrieval-augmented context injection) not instantiated in v1, deferred to v2
  • Test–retest reliability run excluded C4; genuine k-replicate noise floor deferred to v1.1
  • Name-coded counterfactual is a surface proxy (correspondence-audit), not a structural-causal-model intervention; does not satisfy SCM-counterfactual guarantee
  • Proprietary models without shared API access receive trace-only leaderboard entries without held-out private-split verification
  • Per-domain AFB-Score composites are not commensurable across domains (score scales differ in harm magnitude)
  • Explicit-attribute robustness arm (beyond name-coded signals) deferred to future work

Relevance to Harnesses / Meta-Harnesses

AgentFairBench is a purpose-built evaluation harness whose architecture is a direct case study in harness design: it ships a layered profile → scaffold → metrics pipeline with pluggable model adapters, a replay adapter for post-hoc audit without re-querying the model, seeded reproducibility over deterministic components, and a CLI wrapping the full evaluation workflow. The arity-matched null finding is a meta-level harness correctness lesson — a statistical artifact introduced by comparing metrics of different arity — directly analogous to bugs that can appear in any multi-condition evaluation harness that computes aggregate statistics across unequal group sizes. The anti-gaming protocol (contamination canary, held-out split, content hashes, PR-based submission) is a reusable design pattern for live leaderboard harnesses that need to resist benchmark saturation.