Institution profile

Allen Institute for Artificial Intelligence

Academic institutionnorthamerica · us
Official website
Research library331linked papers
Opportunities0open roles
Selected work

Representative Papers

Spurious Rewards: Rethinking Training Signals in RLVR

Jun 12, 2025

Reinforcement learning with verifiable rewards (RLVR) surprisingly improves mathematical reasoning performance even under spurious reward signals (e.g., random, formatting-based, or incorrect labels), yet the underlying mechanism remains unclear. Method: We analyze RLVR’s effect on Qwen2.5-Math and contrast it with Llama3 and OLMo2, employing single-step reinforcement and majority-voting pseudo-rewards on MATH-500. Contribution/Results: We find RLVR activates latent reasoning representations embedded in Qwen2.5-Math’s pretraining—without requiring ground-truth rewards—by specifically amplifying “code-style reasoning,” a critical inference behavior whose frequency rises from 65% to >90%. This phenomenon is model-specific and absent in other architectures. RLVR achieves a 27.1% accuracy gain on MATH-500, approaching the 29.1% gain attained with true rewards. Our work is the first to demonstrate that RLVR enhances generalization via model-specific restructuring of reasoning patterns, establishing a novel paradigm for unsupervised reasoning alignment.

8 citations1 influentialRead paper

Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding

Jan 15, 2026

Existing open-source video-language models are hindered by the scarcity of high-quality training data and insufficient pixel-level grounding capabilities, limiting their performance on complex vision-language tasks. This work proposes Molmo2—the first open-source vision-language model to achieve high-performance video and multi-image understanding with point-driven, pixel-level localization, without relying on any data generated by closed-source models. Molmo2 leverages efficient data packing, message-tree encoding, bidirectional visual token attention, and a novel token weighting strategy, trained on seven newly curated video datasets and two multi-image annotation datasets. The 8B-parameter variant, Molmo2-8B, outperforms existing open-source models on video counting and captioning tasks, significantly surpasses Qwen3-VL in video grounding, and even exceeds Gemini 3 Pro on select metrics.

6 citations2 influentialRead paper

A Little Depth Goes a Long Way: The Expressive Power of Log-Depth Transformers

Mar 05, 2025

This work investigates how logarithmic depth growth—Θ(log n) in input length n—enhances the expressive power of Transformer models, specifically for state tracking and multi-step reasoning over short-to-medium-length sequences. Method: We combine theoretical analysis—including AC⁰ complexity lower bounds, circuit simulation, and attention-based modeling—with empirical validation. Contribution/Results: We establish, for the first time, that Θ(log n) depth suffices for Transformers to recognize regular languages and decide graph connectivity—two classical tasks provably intractable for fixed-depth Transformers. Our results demonstrate that logarithmic depth scaling improves reasoning efficiency more effectively than width expansion or naive layer stacking. Moreover, the theoretically derived minimal depth closely matches empirically observed training requirements, providing a verifiable, theoretically grounded design principle for lightweight sequence reasoning models.

3 citations1 influentialRead paper

Scideator: Human-LLM Scientific Idea Generation Grounded in Research-Paper Facet Recombination

Sep 23, 2024arXiv.org

To address the bottleneck in scientific idea generation, this paper proposes a hybrid active paradigm that decouples and recombines paper-level elements. Methodologically, it implements a four-module RAG architecture: (1) LLM-based automatic extraction of key elements—objectives, mechanisms, and evaluations—from multiple papers; (2) interactive user-driven recombination to synthesize novel research ideas; (3) semantic retrieval for grounding; and (4) an interpretable novelty scoring model enabling closed-loop validation. The core contribution lies in the first integrated synthesis of element-level decomposition, analogy-driven ideation, and iterative novelty assessment. In a controlled study with 19 computer science researchers, the approach significantly increased the output of high-interest ideas compared to a search-engine-plus-LLM baseline (p < 0.01), empirically validating its effectiveness in enhancing both the quality and efficiency of human-AI collaborative scientific ideation.

2 citationsRead paper

SERA: Soft-Verified Efficient Repository Agents

Jan 28, 2026

Existing open-weight code agents struggle to adapt efficiently and cost-effectively to private codebases, limiting their practical utility. This work proposes an efficient training approach based on Soft Verification-guided Generation (SVG) and Supervised Fine-Tuning (SFT): SVG automatically generates thousands of high-quality execution trajectories from a single private codebase, which are then leveraged via SFT to rapidly construct a specialized code agent. To the best of our knowledge, this is the first method to achieve expert-level adaptation to private repositories under fully open-source conditions, attaining state-of-the-art performance while reducing training costs by 26× compared to reinforcement learning and by 57× relative to existing synthetic data approaches.

1 citations1 influentialRead paper
Recent publications

Latest Papers