Gemini-3.1-Pro

2026-11-02

RESEARCH · 1 SOURCE · Apple Machine Learning Research

Online RL with Planning‑Grounding Agent and GiGPPO improves UI navigation on live iOS apps

The paper introduces an online reinforcement learning framework that trains a decoupled Planning agent and UI Grounding model directly on live iOS apps, using an LLM-as-a-Judge to supply dense step-level and terminal rewards and optimizing with PPO plus a new neighbor-conditioned critic (GiGPPO). Evaluations on UINavBench, iOSWorld and AndroidWorld show the authors' 8B RL agents substantially outperform supervised-finetuned and terminal-only RL baselines, achieve ~7× better out-of-distribution generalization to Android, and narrowly beat or match large VLM baselines (outperforming Gemini-3.5-Flash and tying with Gemini-3.1-Pro) on real iOS tasks.

7.0

2026-09-29

RESEARCH · 1 SOURCE · Apple Machine Learning Research

The Communication Bottleneck: round-trip study of tree-structured expression serialization in language models

The paper proposes a round-trip protocol to measure how well tree-structured arithmetic expressions survive serialization into natural-language word problems and back. Evaluating 16 models pairwise, the authors find the channel is lossy and asymmetric (swapping generator and extractor can change accuracy by up to 60.4 points), the best generator–extractor pair achieves 92.9% round-trip correctness, at least 73.6% of failures originate in generation, structure (operator count, depth, right-branching) drives difficulty more than model family, and ~3,600 targeted fine-tuning examples substantially improve open-weight models (raising them above untrained Gemini-3.1-Pro and helping in disjoint-domain cases).

7.0

2026-09-21

RESEARCH · 1 SOURCE · arXiv cs.AI

Round-trip study finds lossy, asymmetric serialization of tree-structured expressions in language models

The arXiv paper (arXiv:2609.21509v1) proposes a round-trip protocol where one model generates a word problem from a procedurally created arithmetic expression and another model extracts the original expression; symbolic equivalence is used as an exact oracle. Evaluating 16 models pairwise, the study finds the natural-language channel is lossy and asymmetric (generation vs extraction can shift accuracy by up to 60.4 points), most failures stem from generation and tree structure (operator count, depth, right-branching) drives difficulty, and roughly 3,600 targeted fine-tuning examples substantially improve open-weight models above an untrained Gemini-3.1-Pro baseline.

7.0

2026-04-24

MODELS · 1 SOURCE · DeepSeek

DeepSeek publishes open-source DeepSeek‑V4 preview with 1M-token context

DeepSeek has open-sourced a preview of DeepSeek‑V4, offering two variants: DeepSeek‑V4‑Pro (1.6T total / 49B active params) and DeepSeek‑V4‑Flash (284B total / 13B active params). The company says both models support a default 1M-token context, introduce novel token-wise compression and DSA (DeepSeek Sparse Attention), claim open-source state‑of‑the‑art agentic and reasoning performance (trailing only Gemini‑3.1‑Pro among models they compare to), and are available now via chat.deepseek.com and updated APIs with compatibility for OpenAI ChatCompletions & Anthropic endpoints; older models deepseek-chat and deepseek-reasoner will be retired on Jul 24, 2026 at 15:59 UTC.

8.0