H Hiring LLM auditPoster PDF ↗

Research companion · November 2026

Same CV.
Different identity.
Different decision?

What supervised fine-tuning changes in LLM CV screening, what further training adds, and what changes inside the model.

Qwen3.5-9B · One shared SFT checkpoint · Three seeds per added stage

One measured example

Same qualifications.
Change only the identity cue.

Sales ManagerCV: 4.5 years in sales
Job: 3+ years required
ModelFemale
she / her
Male
he / him
BaseHireReject
SFTHireHire

Across all 20 gender-cue variants: Base hires 15; SFT hires all 20.

One CV–job pair, greedy decoding. Consistency alone does not establish correctness.
150CVs
450CV–job pairs · 171 jobs
2,700attribute sets
31,050prompts per audit
01

Question & protocol

Hold qualifications fixed.
Vary the identity cue.

Each CV is paired with three jobs. We vary military status, religion or gender through explicit or implicit cues. A set holds the CV, job, attribute family and cue style fixed. It is unstable when it contains both hire and reject decisions.

We compare methods at a matched hire rate by thresholding the hire/reject logit margin. Paired 95% confidence intervals resample candidates, preserving the related sets within each CV.

This development benchmark informed which methods to continue. The results are exploratory. GPT-4o agreement is a reference proxy, not independently labelled hiring quality.

Base → SFT62,761 cue-free reference decisions · one shared checkpoint
DPODirect preference optimization. 4,880 pairs: teacher-written chosen responses and SFT-sampled mistakes.
GRPO + decayGroup relative policy optimization. 4,154 prompts with cue-free decision-reference rewards.
Self-distillationMatch the frozen SFT teacher’s cue-free decision-token probabilities.
LoRA: low-rank adaptation, rank 32. Each added stage starts from SFT and uses seeds 42, 43 and 44.
02

Behavioural results

SFT delivers the main stability gain.

Unstable sets · matched 22.51% hire rate
11.8% → 1.8%

Base → SFT: −9.95 percentage points
Paired 95% CI [−12.07, −7.83] pp

GPT-4o reference agreement78.8% → 82.4%

At each model’s own greedy hire rate.
This comparison uses a different rate convention.

Held-out: SFT’s gain replicates on 300 unseen CVs.

On the frozen RJDB evaluation, unstable sets fall from 24.9% to 2.2% at a matched 36.9% hire rate. The paired change is −22.75 pp, 95% CI [−24.86, −20.61]. Qualification preference remains high: 96.3% → 95.7%, change −0.67 pp [−3.00, +1.67].

No additional consistency gain detected; some DPO/GRPO seeds reduce qualification preference

Qualification preference is the share of triples where the matched CV has a higher hire/reject margin than its weakened counterpart on the cue-free prompt. This is a controlled preference measure, not hiring accuracy.

Added stage / seedConsistency gain vs SFT, pp [95% CI]Qualification preference change, pp [95% CI]
DPO s42+0.39 [−0.56, +1.31]−4.00 [−7.00, −1.00]
DPO s43−0.33 [−1.36, +0.58]−4.67 [−8.00, −1.33]
DPO s44+0.06 [−0.94, +1.01]−1.67 [−4.00, +0.67]
GRPO + decay s42−0.16 [−1.11, +0.83]−2.00 [−4.00, −0.67]
GRPO + decay s43+0.08 [−0.87, +1.03]+0.00 [−1.67, +1.67]
GRPO + decay s44−0.65 [−1.80, +0.44]−4.00 [−6.67, −2.00]
Self-distillation s42+0.23 [−0.73, +1.15]+1.33 [+0.33, +2.67]
Self-distillation s43+0.65 [−0.19, +1.56]+0.67 [−0.67, +2.33]
Self-distillation s44+0.70 [−0.21, +1.60]+0.67 [−1.00, +2.33]

Consistency gain = SFT instability − model instability, so positive is better. All nine consistency intervals include zero at the common 36.9% rate. Intervals are pointwise paired candidate-bootstrap intervals, not adjusted across runs. Four of six DPO/GRPO preference intervals lie below zero; this is a seed-level finding.

The 300 CVs and jobs were unseen, from a source represented in training. The weakened CVs are GPT-4 rewrites and differ in length. The preference interval does not establish equivalence. Checkpoints were fixed before this evaluation; ten jobs used 11.4 allocated GPU-hours.

Instability estimates and intervals (CSV) · Qualification preference (CSV) · Registered protocol, results and caveats ↗

The development DPO drop did not replicate
ComparisonDevelopment, 150 CVsHeld-out, 300 CVs
DPO consistency gain at 35% hire, three-seed range−4.52 to −2.50 pp+0.23 to +1.06 pp
DPO consistency gain over 30–35%, three-seed band-mean range−2.55 to −1.31 pp+0.09 to +0.49 pp
Runs with gain CI above zero: own rate / common point6 of 36 / 1 of 360 of 9 / 0 of 9

The range rows compare the same three DPO checkpoints. The count row describes different sets of evaluated runs and is not a matched-cohort comparison. Common points differ between datasets: 22.51% and 36.88%. Band means and curves are descriptive.

Held-out curves (CSV) · Held-out band means (CSV) · Development band means (CSV) · All development runs (CSV)

Development benchmark: effects vary with hire rate.

Move the slider to inspect descriptive three-seed means on the measured 0.5-point grid. The high-rate DPO drop below did not replicate on the held-out CVs.

Visible methods
10%35%

Change in unstable sets, percentage points. Lower is better. Thin seed trajectories, when shown, are not confidence bands. Curves are descriptive; paired intervals are reported below at 22.51%.

At 22.51% hire: only 1 of 9 seed intervals excludes zero

Mean changes range from −0.285 to −0.165 pp. These are descriptive seed means, without an aggregate three-seed CI.

Added stageSeedChange vs SFT, ppPaired 95% CI, pp
Download comparison table (CSV) ↓
03

Residual instability

Which cases remain unstable
also changes.

For seeds 43/44, unstable-set overlap is much lower than a control that holds SFT scores fixed and recalibrates its threshold to each trained model’s hire rate.

Other seed pairs have different rate patterns. The complete table below includes them, so this comparison is not presented as a universal result across seed pairs.

Jaccard measures the intersection divided by the union of unstable sets. Sets are not distinct people: one CV contributes multiple related sets.

● Actual overlap○ SFT rate control
DPO0.26 → 0.83
GRPO + decay0.28 → 0.92
Self-distillation0.45 → 0.86
Seeds 43/44, each arm at its own greedy hire rate. The control matches hire counts, not numeric thresholds.
All seed pairs and threshold controls
MethodSeedsActual JaccardSFT rate control
Download overlap table (CSV) ↓
04

Inside the model

A smaller cue effect.
A weaker candidate hub.

What activation patching tests

Run a pair with cue A and cue B, swap one internal activation, and measure the resulting change in the cue-dependent decision margin. The analysis identifies components that carry or oppose the cue effect.

L15.H11 · combat / implicit0.66 → 0.13logits removed per pair by patching

Same 64 Base-selected pairs. This selected contrast does not estimate a population-wide effect or validate every edge in a causal circuit.

Selection changes the estimated magnitude.

On the 64 Base-selected gender/implicit pairs, the cue effect falls 72%. Across all 450 pairs, it falls 25.3%, with 95% CI [12.0%, 36.7%].

ContrastSelected 64All 450
Combat / implicit66%51.9%
Combat / explicit77%65.8%
Religion / implicit61%60.9%
Gender / implicit72%25.3%
Mean absolute cue-effect reduction from Base to SFT. Full-grid results and intervals ↗
After SFT: component-effect profiles correlate with SFT, r = 0.65–0.98, for one selected checkpoint per stage. These probes use 64 SFT-selected pairs per contrast, except gender with 45. Correlation does not establish an identical causal pathway.
Development margin scale and a possible interpretation

On the development benchmark, the median margin-scale factor relative to SFT is 2.0–2.7× for the six DPO/GRPO runs and 0.5–0.5× for the three self-distillation runs. These are development-set diagnostics, separate from the held-out qualification control.

Decision-focused training may change margin scale and the ordering of borderline cases without improving every quality measure. This is a hypothesis to test: the scale diagnostic does not establish a cause for the held-out qualification-preference changes.

Margin-readout measurements (CSV) · Scale-control protocol ↗ · Rank-scale decomposition ↗

Parameter-matched window LoRA: a completed follow-up

Restricting LoRA to layers 15–20 at rank 32 uses 11.4M parameters, versus 58.2M for full-layer rank 32. Increasing the window rank to 163 nearly matches the count at 58.1M.

Window armParametersMean gap vs full, pp
Layers 15–20, rank 3211.4M+0.598
Layers 15–20, rank 16358.1M+0.388

Three-seed descriptive means at each full-layer arm’s own hire count, using the original served merged-checkpoint evaluation. Only 1 of 3 matched-budget seed intervals excludes zero. This does not establish equivalence or a general failure of localization-guided editing. Parameter matching also changes rank and optimization geometry.

Per-seed results, protocol and provenance ↗
05

NVIDIA-accelerated audit

Performance, with its measurement basis.

Generation · one GH2007.3×

vLLM vs HF batch 32

470 s → 64.7 s for 3,105 prompts. Generation only, two repeats. HF batch 8: 1,091 s, or 16.9×.

Full audit generation793 s

0.105 kWh per run

31,050 prompts; medians of four runs. Wall time includes loading and writing, excludes scoring. Generation: 45.65 requests/s.

Training · 1 → 4 GH2003.71×

Wall speed at +8% GPU-h

Training: 2,040 s → 550 s. With startup: 2.67× speed at +50% GPU-h. Global batch 32, 40 steps.

Mechanistic audit20 min

42 GiB maximum sampled peak

Median of 22 runs with 64 pairs each, one GH200. Separate full margin scoring: 32 min, 87 GiB peak. Each pass fits without CPU offload.

CSCS Alps · vLLM 0.31 · TRL 1.15 / PEFT · PyTorch DDP, bf16. Energy is GPU-device energy measured by NVML, not whole-node energy. Benchmark ledger and rerun details ↗

06

Read, reproduce, inspect

The evidence behind the poster.

For printing: download the PDF with 1-inch margins. Both versions are 43 × 24 inches.

References and model resources
  1. Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. 2021.
  2. Rafailov et al. Direct Preference Optimization. 2023.
  3. Shao et al. DeepSeekMath, introducing GRPO. 2024.
  4. Hase et al. Does Localization Inform Editing? NeurIPS 2023.
  5. Zhang & Nanda. Towards Best Practices of Activation Patching in Language Models. ICLR 2024.
  6. An et al. Do Large Language Models Discriminate in Hiring Decisions on the Basis of Race, Ethnicity, and Gender? ACL 2024.
  7. Benchmark / reference: AIHiringBiasAnalysis-LLMs; Fairness-in-AI-Recruitment.
  8. Djinni Recruitment Dataset: anonymised candidate profiles.
  9. Kwon et al. Efficient Memory Management for LLM Serving with PagedAttention. SOSP 2023.
  10. Qwen3.5-9B; PEFT; TRL.