Research
Publications by our team.
Publications by team members, written at their universities and labs before or alongside Duffi AI.
-
One Instruction Is a Benchmark: End-to-End Instruction-Following Evaluation with FlexBench
Turns one instruction into its own multi-turn test suite; model rankings and failure modes change from one instruction to the next.
-
Learning When to Think in Latent Visual Space: Reasoning Mode Adaptive Reinforcement Learning
Reinforcement learning that teaches multimodal models when to reason in latent visual space and when in text.
-
To Answer or Not to Answer? Reliability-Oriented Reinforcement Learning for Visual Language Models
A rubric-based reward that teaches vision-language models to abstain on questions they cannot answer, measured on a new benchmark, Reliable-V-Bench.
-
CAMFT: Conflict-Aware Mergeable Fine-Tuning for Large Language Models
Fine-tuning that makes task models easier to merge: each task updates a sparse set of parameters with little conflict across tasks, and merged multi-task models beat those built from standard fine-tuning.
-
From Absolute to Relative: Rethinking Reward Shaping in Group-Based Reinforcement Learning
Relative rankings in place of absolute scores as the reward in group-based reinforcement learning.
-
Think-Augmented Function Calling: Improving LLM Parameter Accuracy Through Embedded Reasoning
A reasoning parameter in each function signature makes the model think before it fills an argument, with no change to the model and full API compatibility.
-
Asymmetric Decision-Making in Online Knowledge Distillation: Unifying Consensus and Divergence
Online distillation in which students follow what teachers agree on, and teachers keep what sets them apart.
-
Large Language Models Know What is Key Visual Entity: An LLM-assisted Multimodal Retrieval for VQA
A language model picks out the visual entities a question needs, so retrieval ignores the rest of the image.