Advancing AI reliability

Research for reliable AI in critical systems

Thoughtworks AI Research Labs conducts research into how AI models can be evaluated, understood and controlled for use in critical environments at scale.

Our research

We focus on rigorous evaluation, interpretability, robustness and model control: how AI behaves, where it fails and how it can be guided. We also explore model interoperability and AI decision-making.

Read the research. Use the code.

Browse the Labs’ projects to access the research behind each one, with links to papers, code and datasets where available. Access the methods, run the code and build on the results.

insights

Post-training NVIDIA Nemotron 3.5 Lightning for enterprise domains

By Gustavo Lujan

Published: August 11, 2026

Using post-training techniques to improve model performance.

insights

EAGLE-3 speculative decoding for NVIDIA Nemotron 3.5 Lightning

By Gustavo Lujan

Published: August 11, 2026

Experimenting with speculative decoding using the EAGLE-3 draft head.

insights

Teaching a model to stop writing like a model

By Allen Roush

Published: August 11, 2026

How can we make an AI model write less like a model and more like a human?

insights

Cultural counterfactuals

By Phillip Howard, Parag Mahajani and Xin Su

Published: July 20, 2026

Discover the 'Cultural Counterfactuals' dataset. We help you uncover hidden bias in AI vision models related to religion, nationality, and socioeconomic status.

Publications and resources

Browse the full set of Thoughtworks AI Research Labs materials, including research papers, technical explainers, open-source code and datasets.

Partners and collaborations

Thoughtworks AI labs sit within a wider network of organizations spanning public AI research, semiconductor innovation, cloud platforms, open source and AI engineering.

These relationships strengthen the lab’s ability to contribute to the methods, tools and technical standards shaping reliable AI.

For partnerships and collaboration inquiries

email ai-labs@thoughtworks.com

Research for reliable AI in critical systems

Our research

Read the research. Use the code.

Post-training NVIDIA Nemotron 3.5 Lightning for enterprise domains

EAGLE-3 speculative decoding for NVIDIA Nemotron 3.5 Lightning

Teaching a model to stop writing like a model

Cultural counterfactuals

Publications and resources

Geometric curriculum coverage for detecting summary incompleteness via Hausdorff distance

Hi-SEMFLOW: Lie algebra–based semantic flow for span-level informal language identification in Hindi

Rethinking skeleton-based action recognition from an action-class prediction distribution perspective

Do tokenizers fail on informal Hindi expressions? Evidence from static, downstream and robustness analyses

TinySQL: A progressive text-to-SQL dataset for mechanistic interpretability research

A semantic parsing framework for end-to-end time normalization

Cultural awareness in vision-language models: A cross-country exploration

MultiFeature graph convolutional network for OCR verification

Retrieval augmented forecasting for generalization

A superpersuasive autonomous policy debating system

Pruning the paradox: How CLIP’s most informative heads enhance performance while amplifying bias

Transformer-based temporal information extraction and application: A review

Antislop: A comprehensive framework for identifying and eliminating repetitive patterns in language models

p-less Sampling: A robust hyperparameter-free approach for LLM decoding

Beyond I am sorry, I can’t: dissecting large language model refusal

Towards transparent AI grading: Entropy as a signal for human-AI disagreement

Investigating the robustness of retrieval-augmented generation at the query level

Lie Algebra based semantic flow for incompleteness detection in summarization

Beyond linear steering: Unified multi-attribute control for language models

Training-free mitigation of language reasoning degradation after multimodal instruction tuning

Learning from reasoning failures via synthetic data generation

Is your paper being reviewed by an LLM? Benchmarking AI text detection in peer review

Scaling knowledge graph construction through synthetic data generation and distillation

Teaching a model to stop writing like a model

EAGLE-3 speculative decoding for NVIDIA Nemotron 3.5 Lightning

Post-training NVIDIA Nemotron 3.5 Lightning for enterprise domains

Cultural counterfactuals

Curveball steering — Geometry-aware non-linear steering to control LLM behavior

Anti-slopping — An innovation for rectifying LLM writing clichés

Concept consistency score

p-less sampling: A robust LLM decoding strategy

Steering smarter

Evaluating LLM-generated summaries using the Lie algebra framework

Distribution-aware feature selection for SAEs

Towards transparent AI grading: Entropy as a signal for human-AI disagreement

The next frontiers in AI — according to industry leaders

Beyond linear steering: Unified multi-attribute control for language models

Calculating uncertainty in generative AI

TinySQL

Evaluating LLMs using semantic entropy

LLM benchmarks, evals and tests

Turning up the heat: Min-p samling for creative and coherent creative outputs

Decoding LLM uncertainties for better predictability

A surprisingly effective way to estimate token importance in LLM prompts

Probabilistic machine learning and weak supervision

A gentle introduction to machine teaching

Teaching a model to stop writing like a model

EAGLE-3 speculative decoding for NVIDIA Nemotron 3.5 Lightning

Post-training NVIDIA Nemotron 3.5 Lightning for enterprise domains

Cultural counterfactuals

Geometric curriculum coverage for detecting summary incompleteness via Hausdorff distance

Curveball steering — Geometry-aware non-linear steering to control LLM behavior

Anti-slopping — An innovation for rectifying LLM writing clichés

Hi-SEMFLOW: Lie algebra–based semantic flow for span-level informal language identification in Hindi

Rethinking skeleton-based action recognition from an action-class prediction distribution perspective

Do tokenizers fail on informal Hindi expressions? Evidence from static, downstream and robustness analyses

TinySQL: A progressive text-to-SQL dataset for mechanistic interpretability research

Concept consistency score

A semantic parsing framework for end-to-end time normalization

p-less sampling: A robust LLM decoding strategy

Cultural awareness in vision-language models: A cross-country exploration

MultiFeature graph convolutional network for OCR verification

Retrieval augmented forecasting for generalization

Steering smarter

A superpersuasive autonomous policy debating system

Pruning the paradox: How CLIP’s most informative heads enhance performance while amplifying bias

Transformer-based temporal information extraction and application: A review

Antislop: A comprehensive framework for identifying and eliminating repetitive patterns in language models

Evaluating LLM-generated summaries using the Lie algebra framework

p-less Sampling: A robust hyperparameter-free approach for LLM decoding

Beyond I am sorry, I can’t: dissecting large language model refusal

Distribution-aware feature selection for SAEs