Manchester researcher bridging mechanistic interpretability and neuro-symbolic computational argumentation
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-14) 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 yet scored: no publication-volume data is in this snapshot. This axis would measure total research output — paper counts, venue distribution, and yearly production — once a fuller bibliographic record becomes available.
Impact—
Not yet scored: citation counts and citation-based influence metrics are absent from the available data. This axis would measure how widely the author's two reviewed papers and broader body of work are cited across the literature.
Focus0/100
The snapshot's field distribution is an exact 50/50 split between mechanistic interpretability of neural networks / AI and computational argumentation / neuro-symbolic NLP, yielding a minimal focus concentration score (0/100). Rather than one core thread, the reviewed 2026 work spans two distinct research programmes — interpretability case studies of transformer world models and prover-verified autoformalization of arguments — each pursued at similar depth.
Collaboration32/100
The recorded collaboration footprint is compact: 4 distinct coauthors appear in the snapshot, giving a collaboration score of 32/100. Pierre Beckmann, Matthieu Queloz, Xin Quan and Reto Gubelmann are the recurring partners, and on both reviewed 2026 papers the author sits in the last position with the corresponding-author flag unset — a pattern consistent with senior, supervisory authorship within a small, stable team.
Activity—
Not yet scored: the data covers only two 2026 preprints, so multi-year productivity trends cannot be computed. This axis would measure publication cadence and sustained research activity across consecutive years.
Reproducibility78/100
Reproducibility is the strongest measured axis: the mean readiness score across the 2 reviewed papers is 78/100. 'World Modeling in Transformers' is rated high readiness, consistent with Evolve's note of a rigorous causal mechanistic case study, while 'Autoformalizing Argumentative Material Inferences' is medium, with faithfulness resting on LLM-generated probes. The pattern suggests consistent attention to verification, from causal tracing to prover-checked proofs.
Research summary
Andre Freitas is a researcher at the University of Manchester whose reviewed work to date sits at the intersection of mechanistic interpretability and computational argumentation / neuro-symbolic NLP — the snapshot splits his catalog evenly (0.5/0.5) between these two fields. In 2026 he appears as last author on two arXiv preprints reviewed in the Evolve catalog. 'World Modeling in Transformers' (2026) traces a TaxiGPT model's failures to superposition interference rather than an incoherent world map — a causal mechanistic case study Evolve rates as rigorous while flagging toy-world, single-run caveats and high reproducibility readiness. 'Autoformalizing Argumentative Material Inferences' (2026) presents GUARD, an LLM-plus-Isabelle pipeline that converts implicit argumentative inferences into prover-verified proofs, improving VFR by 33–35 points, with medium reproducibility readiness and faithfulness resting on LLM-generated probes. The collaboration footprint in this snapshot is compact: 4 distinct coauthors, with Pierre Beckmann, Matthieu Queloz, Xin Quan and Reto Gubelmann recurring. Radar scores reproducibility at 78/100 and collaboration at 32/100, with a minimal focus concentration (0) reflecting the even two-field split; output, impact and activity remain unscored for lack of data.
Collaboration pattern
Collaboration in this snapshot is small and stable: 4 distinct coauthors (score 32/100), with Pierre Beckmann, Matthieu Queloz, Xin Quan and Reto Gubelmann as recurring partners. On both reviewed 2026 papers the author is listed last with the corresponding-author flag unset, suggesting a senior supervisory role within a compact team rather than large consortium work.
Research keywords
Highlights
Tracing transformer world-model failures to superposition interference
GUARD: prover-verified autoformalization of argumentative inferences
Timeline
- 2026Published 'World Modeling in Transformers' on arXiv as last author; Evolve rates it a rigorous mechanistic case study with high reproducibility readiness.
- 2026Published 'Autoformalizing Argumentative Material Inferences' on arXiv as last author; GUARD (LLM + Isabelle) delivers 33–35 VFR points with medium reproducibility readiness.
- 2026Both papers enter the Evolve catalog, anchoring a 50/50 field split between mechanistic interpretability and computational argumentation / neuro-symbolic NLP in this snapshot.
Papers in the Evolve catalog
Frequent co-authors
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.
- Profile is name-resolved from public bibliographic data; ORCID and country are absent from the source record.
- Radar scores derive from a small snapshot (2 reviewed papers, 4 distinct coauthors); output, impact and activity axes are unscored for insufficient data.
- Both catalog papers are 2026 arXiv preprints; no peer-reviewed venue or citation data is in scope.
- Coverage is limited to papers reviewed in the Evolve catalog; other publications by this author are not reflected.