Research program · Trustworthy machine learning
Testing whether machine learning results mean what they claim
Models are judged by numbers: benchmark accuracy, subgroup AUC, internal validation. I test whether those numbers support the decisions built on them, and build the audits and open infrastructure that show when they do not, across LLM evaluation, clinical risk, lending and network security.
- Benchmark Evaluation
- Partial Identification
- Fairness
- External Validation
- Deployment Shift
- Calibration
- Now
- Applying to MS programs in Computer Science and Machine Learning for Fall 2027
- Latest · Oct 2026
- Submitted “Which Leaderboard Orderings Survive a Human Reader?” to DMLR.
Three findings, one program
- Benchmark evaluationSubmitted · DMLR
A human audit certifies fewer than half the orderings on a published LLM leaderboard, and none without its one stated assumption.
174 of 351 orderings certified
Read the paper → - FairnessSubmitted · ML4H 2026 Findings
A subgroup AUC gap narrows while the screening decision flags fewer Black respondents.
542 fewer flagged per 100,000
Read the paper → - Clinical AIPublished
External validation on over 1.28 million records exposes an age-shaped failure that internal validation hides.
AUC 0.794 → 0.717, and 0.607 over 60
Read the paper →
Research vision
A model can top a benchmark, pass validation and clear a fairness audit, and the number can still fail to mean what it claims. My research asks the question evaluation skips: not “is the score high?” but “what does the evidence identify, and does the decision hold for the people it reaches?”
- Identificationcan this ranking be separated at all?
- Scoring ruleswho decided the answer was wrong?
- Decision thresholdsdoes the fairness metric match the decision?
- External validitydoes it survive a new population?
- Deployment disciplinewhen should the gate stay closed?
Flagship research · Submitted · DMLR
Which leaderboard orderings survive a human reader?
A 400-item human audit certifies 174 of 351 pairwise orderings among 27 models under one stated assumption, and none without it.
Swipe to see the whole figure
- 01Read every response three waysthree verifiers, 35,748 published responses, 27 models
- 02Audit a sample by hand400 items read by a human
- 03Bound each pairwise gapone stated assumption, reported as a curve
- 04Certify orderingsall 351 pairs at once, exact in finite samples
The claim
The verifiers disagree on 1,159 responses, and on 891 of them a verifier that gave no credit had found no answer it could read. The audit certifies 174 of 351 orderings at ε = 0.01, 208 at ε = 0, and none without the assumption. When the leaderboard itself changed verifier in February 2025, 15.9% of the model pairs that both scores order strictly reversed (3,366 models, each compared on its own results file before and after, so on the same responses).
- 174 / 351orderings certified at the reference tolerance (ε = 0.01)
- 1,159of 35,748 responses where the three verifiers disagree
- 15.9%of strictly ordered model pairs reversed when the leaderboard changed verifier
The same question, three more settings
- Submitted · ML4H 2026 FindingsHigher AUC, Fewer Cases FlaggedFederated training narrows the White-Black AUC gap while a race-blind screening policy flags fewer Black respondents.
- In preparationConfidently WrongOn a new network, DoS detectors do not decay toward chance: many pass through it and score attacks below benign traffic.
- arXiv preprint 2604.19298IndiaFinBenchA public benchmark of 406 expert-annotated questions on Indian financial regulation, with twelve LLMs evaluated zero-shot.
Research trajectory
Every question was earned by the one before it
- 2025Healthcare AI
Does a model that passes internal validation survive a new population?
Clinical ML meets external validation
Diabetes risk models built on NHANES are tested on 1.28M BRFSS records. External AUC falls 9.7%, and adults over 60 lose the most. First evidence that the benchmark hides the failure; published in IEEE Xplore in May 2026.
- 2025-2026Fairness
How large is a lending disparity at national scale, and where does it live?
Inside the institution: a 42M-application gap
Across 42 million U.S. mortgage applications the Black-White approval gap is 14.95 points, and 74.6% of it sits inside individual lenders rather than between them.
- Apr 2026Benchmark Evaluation
Do LLM benchmarks cover the regulation they would be deployed under?
IndiaFinBench
406 expert-annotated questions on SEBI and RBI regulation, released with code and model outputs. Twelve LLMs score 70.4% to 89.7% zero-shot, in three statistically distinct tiers.
- Aug 2026Network Security
When a detector fails on a new network, is the threshold or the ranking broken?
Confidently wrong detectors
108 of 240 cross-network evaluations score below chance. Eleven of twelve testbed pairs fail at the ranking, where recalibration cannot help.
- Sep 2026Fairness
Can a fairness metric improve while the decision it justifies gets worse?
Higher AUC, fewer cases flagged
Federated screening narrows the White-Black AUC gap from 0.0075 to 0.0005 while the sensitivity gap at fixed capacity widens from 0.009 to 0.034. An exact decomposition shows why the two readings disagree.
- Oct 2026Benchmark Evaluation
Which leaderboard orderings survive when a person reads the responses?
Which orderings survive a human reader?
On the archived MATH-Hard board, a 400-item human audit certifies 174 of 351 pairwise orderings among 27 models under one stated assumption, and none without it.
- NextFuture
Can evaluation report its own uncertainty by default?
Evaluation as infrastructure
Verifier reporting, audit budgets and threshold-aware fairness checks that ship with the number, so a reported result carries the range the evidence supports.
Research map
One ecosystem, not a list of results
Hover or tap a theme to see its papers
Publications & manuscripts
Six studies, one argument
Showing all 6
Which Leaderboard Orderings Survive a Human Reader?
A 400-item human audit certifies 174 of 351 pairwise orderings among 27 models under one stated assumption, and none without it.
- Benchmark Evaluation
- NLP · LLMs
Higher AUC, Fewer Cases Flagged: Subgroup Fairness Metrics Can Reverse at the Decision Threshold in Federated Diabetes Screening
Federated training narrows the White-Black AUC gap while a race-blind screening policy flags fewer Black respondents.
- Fairness
- Healthcare AI
- Privacy · Federated
Confidently Wrong: Ranking Inversion in Cross-Network Denial-of-Service Detection
On a new network, DoS detectors do not decay toward chance: many pass through it and score attacks below benign traffic.
- Network Security
- Deployment Shift
Comprehensive Evaluation of Machine Learning for Type 2 Diabetes Risk Prediction: Large-Scale External Validation and Fairness Analysis
Internally validated diabetes models lose discrimination on 1.28M external records, and adults over 60 lose the most.
- Healthcare AI
- Fairness
- Deployment Shift
IndiaFinBench: An Evaluation Benchmark for Large Language Model Performance on Indian Financial Regulatory Text
A public benchmark of 406 expert-annotated questions on Indian financial regulation, with twelve LLMs evaluated zero-shot.
- Benchmark Evaluation
- NLP · LLMs
- Financial ML
Persistent Racial Disparities in U.S. Mortgage Approval: Evidence from 42 Million Applications, 2020-2024
The public-data Black-White mortgage approval gap at national scale: how large, where it lives, and how it responds to institutional boundaries.
- Fairness
- Causal Inference
Research infrastructure
One reproducible ecosystem
The research ships as usable infrastructure. Where a paper’s code is public, it regenerates the reported numbers from committed result files, and the datasets, benchmarks and audit libraries below are documented for other people to reuse.
- Code · Results · Leaderboard
leaderboard-orderings
Code, human-audit data and results for certifying which orderings on the MATH-Hard board of the Open LLM Leaderboard survive a human reader.
- Benchmark · Reproduction
trustshift
Four real deployment shifts with committed prediction files, so every reported number regenerates without the raw third-party data.
doi:10.5281/zenodo.21739501 - Dataset · Live demo
IndiaFinBench
406 expert-annotated regulatory questions, per-item judge verdicts, model predictions, and a deployed retrieval demo.
doi:10.5281/zenodo.21739533 - Python library · PyPI
fairscope
Subgroup-stratified, calibration-aware fairness auditing: DeLong CIs per subgroup, subgroup ECE/MCE, gap significance testing, recalibration, with healthcare, NLP, federated and lending modules.
pip install fairscope - Code · Corrected results
Federated-Diabetes
Ten-seed federated screening pipeline, with the defects found in an earlier version documented alongside the fixes.
- Runtime audit
aria-audit
Five-axis runtime audit for locally deployed LLMs: calibration, faithfulness, consistency, equity, attribution, wrapping any model on consumer hardware.
- Replication code · Lending
hmda-racial-disparities
The pipeline that regenerates the mortgage-disparity results from their result files.
Research output
- 6papers: 1 published, 3 submitted or under review, 1 preprint
- 351leaderboard orderings tested against a human audit
- 42Mmortgage applications analysed
- 1.28Mrecords in external clinical validation
- 406expert-annotated benchmark questions
- 3public datasets on Hugging Face
The next five years
Questions that remain open
- Q1
Can every reported benchmark number name the verifier behind it?
The leaderboard paper shows that a human audit certifies fewer than half of a published board’s orderings. The next step is making the verifier, its count of unreadable responses and an audited range default outputs of evaluation harnesses, with audit labels spent where they narrow comparisons most.
- Q2
Can fairness be audited at the decision instead of the metric?
Subgroup AUC reads within-group ranking, while a screening threshold acts across groups. Audits should report who a deployed threshold actually flags, at the capacities a programme really runs, with intervals on the gap.
- Q3
Can deployment failure be diagnosed before labels arrive?
Detectors moved to a new network can rank attacks below benign traffic while staying confident. The open problem is which label-free signals reveal that failure on the new network, with what guarantees, and how an audit should report inconclusive when they cannot.
News
Where the work stands
Submitted “Which Leaderboard Orderings Survive a Human Reader?” to DMLR.
Submitted the subgroup fairness reversal paper (federated diabetes screening) to the ML4H 2026 Findings track.
Released the flipbudget results dataset and leaderboard on Hugging Face.
Reviewed for W-NUT 2026, the EMNLP 2026 workshop on noisy user-generated text.
Released the TrustShift benchmark and code on GitHub and Hugging Face.
Diabetes external-validation paper published in IEEE Xplore.
Presented the diabetes external-validation paper at CIPHER-2026, NIT Jalandhar.
About
I am Rajveer Singh Pall, an AI researcher studying whether machine learning results mean what they claim, from scoring rules in LLM benchmarks to fairness at the decision threshold and clinical external validation. 6 papers so far: 1 published, 3 submitted or under review. I am applying to MS programs in Computer Science and Machine Learning for Fall 2027.
- Education
- B.Tech, Computer Science and Business Systems
Gyan Ganga Institute of Technology and Sciences, Jabalpur, India · Expected 2027 - Service
- Program committee reviewer, FinNLP 2026Reviewer, W-NUT 2026, the EMNLP 2026 workshop on noisy user-generated text