Antonio Mari

—

Unclaimed · AI-structured profileOpenAlex

Antonio Mari — early-career LLM researcher in agentic benchmark auditing and subliminal-learning mitigation

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Not scored. OpenAlex reports an h-index and i10-index of 0 with per-year counts covering only 2025–2026, so production volume cannot be reliably estimated. This axis would measure total works, preprint throughput, and productivity trajectory.

Impact—

Not scored. No citation record of sufficient depth exists for this identity (h-index 0, i10-index 0). This axis would measure citation uptake, influential works, and the scholarly attention accruing to the 2025–2026 outputs.

Focus0/100

Focus scores 0/100: the OpenAlex field distribution splits exactly 50/50 between LLM agent benchmark evaluation and large language model research, while the topic list spans five clusters — Topic Modeling (3 counts) plus single appearances of Advanced Neural Network Applications, Software Engineering Research, Structural Behavior of Reinforced Concrete, and GANs and Image Synthesis. Such extreme dispersion may partly reflect identity mixing in the record rather than genuine research breadth.

Collaboration50/100

Collaboration scores 50/100, based on 11 distinct coauthors across 4 works in the recent 3-year window. In the two Evolve-reviewed papers Mari holds a middle author position on the Terminal-Bench audit and the last position on the subliminal-learning paper, with no corresponding-author role recorded — a pattern of contributing inside small, ad hoc teams in both supporting and senior slots.

Activity47/100

Activity scores 47/100: OpenAlex indexes 4 works over the last 3 years, with counts by year of 1 in 2025 rising to 3 in 2026. The trajectory is short but clearly upward, concentrated in 2026 arXiv preprints on LLM evaluation auditing and training-time safety; the window is too brief to establish sustained productivity.

Reproducibility38/100

Reproducibility scores 38/100 (reproMean across 2 reviewed papers). The Terminal-Bench audit rates low — its certified labels rest on LLM judges and single oracle runs — while the subliminal-learning mitigation reaches medium readiness through LoRA experiments on 5 open LLMs, bounded by shared-initialization settings.

Research summary

Antonio Mari is an early-stage researcher working on large language models, profiled here from public bibliographic data (OpenAlex) because no institutional affiliation, country, or ORCID was available. The record is short but pointed: 4 works indexed in the last three years (1 in 2025, 3 in 2026), 11 distinct coauthors, an h-index and i10-index of 0, and a field profile split evenly between LLM agent benchmark evaluation and large language model research. Two 2026 arXiv preprints reviewed by Evolve anchor the picture. The first, a middle-author contribution, audits Terminal-Bench tasks whose claimed all-fail results mask fake hardness, splitting 78 certified-unsolved from 47 defective tasks with an adjudication protocol Evolve rates credible, though its labels rest on LLM judges and single oracle runs. The second, a last-author contribution, introduces annealed early-KL fine-tuning that curbs subliminal trait transfer across 5 open LLMs while preserving task gains, with the caveat that evidence is limited to LoRA and shared-initialization settings. Bibliometric scores place collaboration at 50, activity at 47, and reproducibility at 38; output and impact cannot be scored yet. The breadth of topics in the underlying record means this profile should be read with identity-resolution caution.

Collaboration pattern

Collaboration scores 50/100, built on 11 distinct coauthors across 4 works in the recent 3-year window. The Evolve catalog places Mari in a middle position on the Terminal-Bench audit and in the last position on the subliminal-learning paper, with no corresponding-author role recorded. Named repeat collaborators include Edward Lue Chee Lip, Boden Moraski, Tim Knappe, Lang Xiong, Sarvesh Gharat, and Kevin Zhu, suggesting small, rotating teams rather than a fixed lab group.

Research keywords

large language modelsagentic benchmarksTerminal-Benchbenchmark auditingtask validityreward hackingsubliminal learningknowledge distillationKL regularizationLoRA fine-tuningtrait transferchain-of-thought distillation

Highlights

Auditing Terminal-Bench hardness: 78 certified-unsolved vs 47 defective tasks
An Evolve-reviewed 2026 arXiv paper (middle-author position) separates genuine hardness from fake-hardness among claims-all-fail TB3 tasks, yielding 78 certified-unsolved and 47 defective tasks. Evolve rates the adjudication protocol credible but notes the labels rest on LLM judges and single oracle runs; reproducibility readiness is rated low.
Annealed early-KL fine-tuning curbs subliminal trait transfer
A last-author 2026 arXiv contribution reviewed by Evolve: annealed early-KL fine-tuning reduces subliminal trait transfer in 5 open LLMs while preserving task gains. Evidence is rated solid but bounded by LoRA and shared-initialization settings; reproducibility readiness is medium.
A twin focus on evaluation integrity and training-time safety
Across the two reviewed papers, Mari's work clusters at two complementary pressure points in LLM research: auditing whether benchmark hardness is genuine, and preventing unintended trait transfer during distillation-based training — both targeting failure modes that standard metrics miss.

Timeline

2025OpenAlex record opens with 1 indexed work in 2025, before the 2026 burst of arXiv preprints.
2026Middle-author contribution to the Terminal-Bench hardness audit, separating 78 certified-unsolved from 47 defective TB3 tasks (arXiv, Evolve-reviewed).
2026Last-author contribution on mitigating subliminal learning: annealed early-KL fine-tuning validated on 5 open LLMs (arXiv, Evolve-reviewed).

Papers in the Evolve catalog

Frequent co-authors

Edward Lue Chee Lip×1 Boden Moraski×1 Tim Knappe×1 Lang Xiong×1 Sarvesh Gharat×1 Ivan Bercovich×1 Atsushi Yanagisawa×1 Brendan Gho×1 Rajendran Ramesh Babu Manoj Narender×1 Kevin Zhu×1 Madhur Panwar×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric indicators (h-index, i10-index, counts by year, topics, coauthors) derive from a single OpenAlex identity; no ORCID or institutional affiliation was available for cross-verification.
'Antonio Mari' is a relatively common name, and the topic spread — reinforced concrete alongside LLM work — suggests possible identity mixing within the OpenAlex cluster.
Citation-impact data is absent (h-index 0, i10-index 0), so the output and impact dimensions remain unscored.
Only 2 of the indexed works carry full Evolve reviews; the remaining 2025–2026 outputs are not individually assessed.