University of Bucharest NLP researcher on calibrated honesty, provenance and belief revision in language models
2
papers here
2
reviewed by Evolve
2
first-author
2
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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—
No output score yet: the snapshot's raw metrics list h-index 0 and i10-index 0 with only one 2026 work, so publication volume cannot be assessed. This axis would measure total works, venue spread, and the output trajectory over time.
Impact—
Impact cannot be scored: no citation evidence exists in the snapshot — h-index and i10-index are both 0 on a single 2026 work. This axis would measure citations, citing works, and influence on follow-up research.
Focus0/100
Focus scores 0 because attention is evenly split: the field distribution divides 0.5/0.5 between trustworthy, calibrated language-model behavior and conversational AI, and all three recorded topics (AI in healthcare and education, scientific computing and data management, XAI) carry a count of 1 each. With only two papers, no dominant thread has emerged, though both works sit inside NLP.
Collaboration0/100
The collaboration score is 0 because the OpenAlex-derived snapshot records 0 distinct co-authors and no frequent collaborators; both cataloged 2026 papers list him as first and corresponding author, consistent with a currently solo-led pattern. The record covers one publication year and two papers, so this reflects a small sample rather than a confirmed preference for working alone.
Activity20/100
Activity scores 20, consistent with a nascent record: the snapshot counts 1 work in the last three years, and counts-by-year shows a single 2026 entry. Both cataloged papers appeared on arXiv in 2026, so publication cadence cannot yet be judged over a longer window; the low score reflects an early-stage record rather than a documented slowdown.
Reproducibility100/100
Reproducibility scores 100, the maximum, computed across 2 reviewed papers with a reproducibility mean of 100, and both papers carry a high repro-readiness rating. The transparency extends to results: Evolve's reviews record pre-specified checks that fail in both papers, reported openly. The evidence base is small, but the current pattern pairs open artifacts with honest negative reporting.
Research summary
Sebastian Cochinescu is a researcher at the University of Bucharest working in natural language processing, with the snapshot's field distribution splitting evenly (0.5/0.5) between trustworthy, calibrated language-model behavior and conversational AI. The Evolve catalog holds two 2026 arXiv papers, on both of which he is first and corresponding author. The first builds an honesty behavior layer — expressed doubt, provenance-aware assertion, and belief revision — over a fixed 0.5B-parameter LLM; Evolve's review notes that several pre-specified checks fail and that key gains rest on just 60 facts. The second is a registered test of a stateful selection layer for "structured unpredictability" in conversational AI, whose central criteria are honestly reported as failing, leaving a partly tautological lexical-novelty effect. Bibliometric coverage is thin: the snapshot lists 1 work in the last three years, counts-by-year shows a single 2026 entry, and h-index and i10-index are both 0, so output and impact cannot be scored yet. What the record does show is distinctive: both papers rate high on reproducibility readiness, the reproducibility mean across 2 reviewed papers is 100, and failing pre-specified checks are reported openly — a falsification-first working style.
Collaboration pattern
OpenAlex-derived metrics record 0 distinct co-authors, and the Evolve catalog lists no frequent collaborators. On both 2026 papers he is first and corresponding author, so the visible pattern is solo-led work in which he carries the full authorship and communication role. The record spans one publication year and two papers, so this is best read as an early, unpartnered stage rather than an established solitary style; future co-authorships would quickly change this picture.
Research keywords
Highlights
An honesty behavior layer — doubt, provenance, belief revision — on a fixed 0.5B LLM
Registered test of structured unpredictability reports its own failure
Falsification-first publishing with a perfect reproducibility score
Both 2026 papers rate high on reproducibility readiness, giving a reproducibility mean of 100 across 2 reviewed papers; negative results from pre-specified checks are documented openly rather than hidden, defining the profile's working style.
Timeline
- 2026Publishes the arXiv preprint "Truth for Believable AI: Expressed Doubt, Provenance, and Belief Revision as an Engineerable Stance" as first and corresponding author — the first paper in the Evolve catalog.
- 2026Publishes the arXiv preprint "Entropy in Conversational AI: Structured Unpredictability as Inferrable Interiority", again as first and corresponding author — a registered test whose central criteria are reported as failing.
- 2026Snapshot state: counts-by-year records a single 2026 work, 1 work over the last three years, and h-index 0 — an early-stage record with both papers rated high for reproducibility readiness.
Papers in the Evolve catalog
- first authorcorresponding
Truth for Believable AI: Expressed Doubt, Provenance, and Belief Revision as an Engineerable Stance
- first authorcorresponding
Entropy in Conversational AI: Structured Unpredictability as Inferrable Interiority
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 signals (h-index, i10-index, distinct co-authors, counts-by-year) come from an OpenAlex-derived snapshot; identity is resolved via OpenAlex and no ORCID is linked.
- The catalog covers only 2 papers, both 2026 arXiv preprints; no peer-reviewed venue or citation data is available, so output and impact are left unscored.
- Topic labels (e.g. AI in healthcare and education, scientific computing and data management) come from automated topic assignment with one paper each and may overstate topical breadth.
- The reproducibility score of 100 is computed over only 2 reviewed papers and should not be read as a stable long-term track record.