Tongliang Liu

The University of Sydney

Unclaimed · AI-structured profileOpenAlex

University of Sydney researcher in LLM alignment and multimodal AI safety — CoT monitorability and human-likeness.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) 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 publication-volume metrics are available in the current snapshot, so output is unscored. This axis would measure total publication count, output by venue and year, and first/corresponding-author share across the researcher's full record.

Impact—

Citation-based influence metrics are not yet available for this profile, so impact is unscored. The axis would measure citation counts, h-index, and downstream uptake of the alignment and interpretability work reviewed here.

Focus0/100

The focus score is 0, consistent with an even 50/50 split in the reviewed output between machine learning / LLM alignment and AI safety / LLM interpretability (multimodal). The two 2026 arXiv papers address distinct questions: mechanistic causes of chain-of-thought obfuscation in RL-trained vision-language models, and whether RLHF/DPO preference alignment keeps models human-like. The axis reads as a deliberate two-pronged agenda within one safety-and-alignment program rather than scattered work.

Collaboration54/100

The collaboration score is 54, based on 14 distinct coauthors in the reviewed record, with a recurring pool of twelve frequently listed collaborators including Suqin Yuan, Runqi Lin, Muyang Li, Jindong Gu, Lei Feng, and Chris Russell. Authorship sits in middle position on the vision-language interpretability paper and last position on the alignment paper, with no corresponding-author credit recorded — a senior but supporting role within multi-author teams.

Activity—

Publication-cadence data over time is insufficient to score activity. This axis would measure annual output rhythm, recency of preprints, and sustained research pace across recent years, rather than the two 2026 papers alone.

Reproducibility55/100

The reproducibility score is 55, computed from a mean readiness of 55 across 2 papers assessed, both rated medium. The grounds differ: the TAME chain-of-thought study presents strong mechanistic evidence but judge-centric, single-seed evaluation, while the preference-alignment study offers solid 7-8B model evidence whose frontier-scale claims outrun its tested scope.

Research summary

Tongliang Liu is a researcher at The University of Sydney whose recent work, as catalogued by Evolve, sits at the intersection of large language model alignment and multimodal AI safety. The reviewed record comprises two 2026 arXiv preprints split evenly between the two fields (0.5 each in the field distribution): 'Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation Enforcement' and 'Are Human-Aligned Models Models of Humans? A Turing-Test Gap in Preference Alignment'. The first traces chain-of-thought obfuscation in RL-trained vision-language models to SAE-localizable causes and proposes TAME, an activation-enforcement method that restores monitorability; Evolve credits it with a strong evidence chain while noting judge-centric, single-seed evaluation. The second proves that preference alignment via RLHF and DPO shifts models away from human response behavior unless a restrictive reward condition holds; evidence on 7-8B models is solid, though frontier-scale claims outrun the tested scope. Across the two papers the researcher holds middle and last authorship without corresponding credit, working alongside 14 distinct coauthors including recurring collaborators such as Suqin Yuan, Runqi Lin, Lei Feng, Chris Russell, and Jindong Gu. Bibliometric coverage remains narrow: the output, impact, and activity axes lack data, and both scored dimensions (collaboration 54, reproducibility 55) rest on small samples. This profile should be read as an early-stage, name-resolved snapshot built from public bibliographic data and Evolve's paper reviews.

Collaboration pattern

The reviewed record shows 14 distinct coauthors, with a recurrent core of twelve frequently listed collaborators including Suqin Yuan, Runqi Lin, Muyang Li, Jindong Gu, Lei Feng, and Chris Russell. Authorship positions are middle on the vision-language interpretability paper and last on the alignment paper, with no corresponding-author credit in either. The pattern suggests contribution within medium-to-large alignment and interpretability teams, bridging the two halves of the field distribution.

Research keywords

chain-of-thought monitorabilityCoT obfuscationreinforcement learningvision-language modelssparse autoencodersmechanistic interpretabilityRLHFpreference alignmentDPOhuman-likenessTuring testhuman response distribution

Highlights

Localizing and reversing chain-of-thought obfuscation in vision-language models
Evolve's review finds the paper argues that chain-of-thought obfuscation in RL-trained vision-language models has sparse-autoencoder-localizable causes, and that the proposed TAME method restores monitorability through activation enforcement; the evidence chain is rated strong, with the caveat that evaluation is judge-centric and single-seed.
A Turing-test gap: preference alignment drifts from human behavior
Evolve's review finds the paper proves that preference alignment shifts models away from human behavior unless a restrictive reward condition holds, with solid evidence on 7-8B models while frontier-scale claims outrun the tested scope; the gap is framed through a Turing-test lens on human-likeness and human response distributions.

Timeline

2026Publishes (middle author) 'Taming CoT Obfuscation in VLMs: From Mechanistic Evidence to Activation Enforcement' on arXiv, proposing TAME to restore chain-of-thought monitorability in RL-trained vision-language models.
2026Publishes (last author) 'Are Human-Aligned Models Models of Humans? A Turing-Test Gap in Preference Alignment' on arXiv, proving that preference alignment can move models away from human behavior absent a restrictive reward condition.
2026Evolve catalog records both preprints, with the field distribution split evenly between machine learning / LLM alignment and multimodal AI safety / interpretability, and both papers rated medium for reproducibility readiness.

Papers in the Evolve catalog

Frequent co-authors

Suqin Yuan×1 Runqi Lin×1 Muyang Li×1 Guanzhe Hong×1 Jindong Gu×1 Lei Feng×1 Chris Russell×1 Xutao Mao×1 Jianing Zhu×1 Jinman Zhao×1 Xiaowen Chu×1 Cong Wang×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.

Dimension scores derive from a bibliometric snapshot; the output, impact, and activity axes lack sufficient data and are intentionally unscored.
Only the two Evolve-reviewed 2026 arXiv preprints are reflected here; the researcher's wider publication record is not included.
ORCID and country are missing, so the profile is name-resolved from public bibliographic data and may merge or miss researchers sharing the name.
The reproducibility (55) and collaboration (54) scores are computed from small samples (2 assessed papers; 14 distinct coauthors) and should be read as indicative only.