Furu Wei · page 3
Papers
- Examining False Positives under Inference Scaling for Mathematical Reasoning (2025 · 9 citations)
Most math benchmarks for language models award credit by matching the final answer, usually through heuristic rules, and never check the derivation. - ALYMPICS: LLM Agents Meet Game Theory (2025 · 7 citations)
Empirical game theory usually needs experiments with human participants, which are costly, slow and ethically demanding. - ShifCon: Enhancing Non-Dominant Language Capabilities with a Shift-based Multilingual Contrastive Framework (2025 · 6 citations)
Multilingual LLMs fine-tuned on translated data (MSFT) still lag behind in non-dominant languages compared with English. Producing more high-quality translated supervision is costly and error-prone. ShifCon addresses this by changing how the model processes these languages internally rather than by adding data. - Rectified Sparse Attention for Efficient Long-Sequence Generation (2026 · 1 citation)
Large language models that produce very long outputs, such as extended reasoning chains, are slowed down by having to read the entire KV cache at every decoding step. - Towards Stable and Effective Reinforcement Learning for Mixture-of-Experts (2026 · 1 citation)
The paper addresses why reinforcement learning with verifiable rewards (RLVR) often breaks down when the policy is a Mixture-of-Experts (MoE) model, frequently ending in sudden reward collapse. - Learning to Refine: Self-Refinement of Parallel Reasoning in LLMs (2026)
Common test-time scaling strategies, such as majority voting and Best-of-N, can only pick from the candidates a model produces. - Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM Hallucinations (2026)
Probing classifiers trained on LLM hidden states can detect hallucinations, but it has been unclear where these truthfulness signals come from.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/furu-wei/page/3/