Each Reward to Its Relevant Tokens: Reward-Specific Token-Level Credit Assignment for Multi-Reward RL

Project Page  | Code

Featured on 🤗 Hugging Face Daily Papers (Oct 16, 2026)

GDPO normalizes each reward separately but still spreads every reward’s advantage evenly over the whole response, although each verifier checks only a few positions. MRTPO replays the policy’s own response with and without each reward’s verifier information and uses the per-token Jensen–Shannon distance to assign that reward’s advantage to its relevant tokens, without auxiliary models or token-level annotations. It has the best average performance on text-to-SQL, instruction following and tool calling across all six models, and after 10 steps produces the required tool-call format on 97% of BFCL cases, compared with 2% for GDPO.

Role: First Author.

Make Sparse Rewards Count: Density-Aware Reward Aggregation for Multi-Reward RL

Project Page  | Code

Featured on 🤗 Hugging Face Daily Papers (Oct 2, 2026)

We show that under reward-wise normalization, a reward’s advantage energy scales with its active-group density, leaving sparse objectives with weaker learning signals. DARA derives batch-adaptive density weights to balance reward contributions without hand-tuned weights. It reaches high format compliance in up to 26% fewer steps on tool calling and near-saturated length compliance in up to 65% fewer steps on mathematical reasoning, with competitive final performance.

Role: Co-first Author.

ThinkPrior: Zero-Rollout Difficulty Priors for Cold-Start Prompt Selection in RLVR

Project Page

We use an external anchor model to initialize difficulty priors before the first target-policy rollout, then select prompts by expected learnability and update from training outcomes. On Qwen2.5-Math-7B across sixteen seeds, ThinkPrior more than halves early zero-advantage groups and reduces wasted rollouts through step 30 by nearly a fifth, with no detected difference in final accuracy. Combined with DAPO, it reduces generated rollouts by 10.6% at the same update budget.

Role: Collaborating Author.

Abstain-R1: Calibrated Abstention and Post-Refusal Clarification via Verifiable RL

Model  | Dataset

Reinforcement fine-tuning sharpens LLM reasoning, but also pushes models to guess on unanswerable queries. We propose a clarification-aware RLVR reward that jointly optimizes explicit abstention and semantically aligned post-refusal clarification, treating “what is missing” as a first-class training target. The resulting 3B Abstain-R1 model substantially improves abstention and clarification on unanswerable queries while preserving strong performance on answerable ones, matching larger systems like DeepSeek-R1 across Abstain-Test, Abstain-QA, and SelfAware.

Role: First Author.

Supported by: Lambda and CloudRift research grants ($1k compute credits each).

Initiated as a course project for CSCI 5541 (Natural Language Processing) at UMN.

The Blind Spot of Agent Safety: How Benign User Instructions Expose Critical Vulnerabilities in Computer-Use Agents

Project Page  | Code  | Dataset

Featured on 🤗 Hugging Face Daily Papers (Apr 15, 2026)

OS-Blind was used for safety evaluation in the UI-Venus-2 Technical Report.

Computer-use agents can autonomously drive real digital environments, and when misled they can automate harm at scale. Existing safety evaluations focus on explicit misuse or prompt injection; we study a subtler setting where user instructions are entirely benign and harm emerges from the task context or execution. We introduce OS-Blind, a 300-task human-crafted benchmark spanning 12 categories, 8 applications, and 2 threat clusters. Most CUAs exceed 90% attack success rate; even safety-aligned Claude 4.5 Sonnet reaches 73.0%, climbing to 92.7% in multi-agent deployments, showing that alignment mostly fires in the first few steps and rarely re-engages during execution.

Role: Co-first Author.

AtomTree: A Hierarchical Framework for State-Aware Embodied Instruction Following with LLMs

Project Page

We introduce a state-aware hierarchical framework that decides at each step whether to decompose a subgoal or execute a grounded action. Lightweight memory tracks observations and action outcomes, while hierarchical backtracking regenerates subplans when execution stalls. Evaluated zero-shot on ALFRED, AtomTree couples planning with the evolving environment to improve long-horizon embodied instruction following.

Role: First Author.

Multi-Cache enhanced Prototype Learning for Test-Time Generalization of Vision-Language Models

Project Page  | Code

We observed that cache-based test-time adaptation performance is positively correlated with intra-class compactness. To address the unreliability of low-entropy samples under distribution shifts, we propose MCP, which uses an entropy cache for prototype initialization, an align cache to fuse visual and textual information and tighten intra-class distributions, and a negative cache to calibrate high-entropy predictions. We further extend this into the MCP++ framework by introducing cross-modal prototype alignment and residual learning, achieving state-of-the-art generalization on 15 downstream tasks.

Role: Co-first Author.

Mitigating Cache Noise in Test-Time Adaptation for Large Vision-Language Models

Also accepted at ICLR 2025 FM-Wild Workshop

We analyzed the root causes of the performance gap between zero-shot and few-shot TTA, identifying noisy cache labels as a critical bottleneck. We then propose the CRG framework, which maintains positive and negative visual prototypes alongside text prototypes, employs learnable residuals to align modalities, and leverages Gaussian Discriminant Analysis to dynamically model class distributions and suppress noisy samples. Finally, by jointly minimizing prediction entropy and maximizing inter-prototype distances, CRG achieves superior robustness and generalization across 13 benchmarks..

Role: First Author.

FAST-CAD: A Fairness-Aware Framework for Non-Contact Stroke Diagnosis

Project Page  | MIT Tech Review

Stroke is an acute cerebrovascular disease, so we propose FAST-CAD, a DAT + Group-DRO framework that jointly enforces demographic-invariant representations and worst-group robustness for non-contact stroke diagnosis. Built on a 12-subgroup multimodal dataset, it couples adversarial domain discrimination with self-supervised encoders and optimizes worst-group risk, delivering 91.2% AUC and tight fairness bounds backed by domain adaptation and minimax theory.

Role: Collaborating Author.

“Can We Make Feature-3DGS Faster, Better, and Smaller?”

We accelerate and shrink Feature-3DGS with semantic-aware Gaussian pruning and consistency loss, keeping fine details while boosting FPS and mIoU across Replica and Gopher/LindHall scenes.