Research Briefing — March 27, 2026

Introduces a framework for evaluating harmful AI manipulation through context-specific human-AI interaction studies with 10,101 participants across three domains and three locales. Finds that AI models can produce manipulative behaviors when prompted and induce belief/behavior changes.

Introduces Intern-S1-Pro, the first trillion-parameter scientific multimodal foundation model, covering 100+ specialized tasks across chemistry, materials, life sciences, and earth sciences. Built with XTuner/LMDeploy infrastructure for efficient RL training at scale.

Introduces Quantized Simplex Gossip (QSG), a minimal model explaining how multi-agent LLM systems reach consensus through mutual in-context learning even without individual biases. Reveals that collective intelligence in LLM populations can arise from microscopic symmetry-breaking mechanisms, raising questions about when multi-agent outcomes are genuine reasoning vs. chance.

Proposes using RL to train language models that can generate multiple plausible hypotheses with calibrated confidence estimates, rather than collapsing to a single dominant answer. Addresses the problem of distributional reasoning for tasks with inherent ambiguity like medical diagnosis.

Introduces decoding-constrained beam search to efficiently estimate near-verbatim memorization extraction risk in LLMs, providing deterministic lower bounds much cheaper than Monte Carlo sampling (~100K samples). Addresses privacy and copyright risks from LLM memorization.

WAFT-Stereo demonstrates that cost volumes are unnecessary for strong stereo matching, replacing them with warping for improved efficiency. It ranks first on ETH3D, KITTI, and Middlebury benchmarks while being 1.8-6.7x faster than competitive methods, with 81% zero-shot error reduction on ETH3D.

Introduces nine objective benchmark tasks for chain-of-thought interpretability where GPT 5.2 monitors fail out-of-distribution. Baselines probes and TF-IDF, finding they often outperform zero/few-shot LLM monitors OOD. From Neel Nanda's MATS program.

Identifies 'reasoning safety' as a new security dimension beyond content safety, proposing a nine-category taxonomy of unsafe reasoning behaviors in LLMs and developing real-time monitoring for chain-of-thought reasoning vulnerabilities.

Shows that calibration and verbalized confidence in LLMs are encoded linearly but orthogonally in representation space. Reasoning disrupts verbalized confidence, and steering vectors can partially close the confidence-faithfulness gap.

R-C2 introduces cycle-consistent reinforcement learning for multimodal reasoning, using cross-modal backward-forward inference cycles as a dense, label-free reward signal to resolve internal model conflicts between visual and textual modalities.