搜索“RAG”

able judgment. To train this behavior, we first perform SFT + cold-start initialization to establish stable tool-use patterns. We then apply a two-stage GRPO process: Stage 1 encourages correct tool selection, while Stage 2 refines reward accuracy using verifiable rewards that balance correctness and tool efficiency. Together, these components equip ARM-Thinker with reliable, interpretable, and tool-augmented multimodal reasoning capabilities. 💡 Highlights 🔥 Agentic "Think–Act–Verify" Paradigm A

匹配字段正文

know how Marble can supercharge their creativity and productivity by providing the power of world‑building at their fingertips. Some of our early users are already giving us encouraging feedback, and we're eager to hear from more of you as we continue to improve our technology and product. What used to take weeks and months, I can now do in minutes, and have full control of being in a 3d space. [...] integrating Marble's Image to 3D World (spz, ply) results into my project for games and interac

匹配字段正文

at is rendered as an oriented quad (technically two triangles) enclosing the splat's projected 2D ellipse. The four vertices are calculated using a custom vertex shader, then the fragment shader computes the opacity for each pixel by evaluating the Gaussian profile, and finally the GPU hardware blends the color into the frame buffer. Scaling up to huge worlds Since Spark's initial launch, 3DGS has been growing in popularity but also in scale. 3D scans today often exceed tens of millions of splat

匹配字段正文

a significant portion of checkpoints before real-world testing, and reserve costly hardware evaluation for the most promising policies. The result is faster iteration, broader coverage, and substantially lower cost. A useful simulation need not match real-world success rates exactly. It must support the same decisions as reality: identify where policies succeed and fail, rank which policies are better, and predict whether improvements during training will carry over to hardware. Below, we use an

匹配字段正文

r example, it generates the back side of the robot, and it guesses that there should be a grassy lawn next to the pool. From a single image, Atlas generates views from any angle. Drag to change the view. Generating with Spatial Context Similar to an LLM, Atlas first encodes its inputs into a context, then generates outputs conditioned on the context. However, Atlas is unique because each image is grounded at a 3D position in space; this forms a spatial context. Managing this spatial context unlo

匹配字段正文

iants: FireRedASR-LLM: Designed to achieve state-of-the-art (SOTA) performance and to enable seamless end-to-end speech interaction. It adopts an Encoder-Adapter-LLM framework leveraging large language model (LLM) capabilities; FireRedASR-AED: Designed to balance high performance and computational efficiency and to serve as an effective speech representation module in LLM-based speech models. It utilizes an Attention-based Encoder-Decoder (AED) architecture Evaluation Results are reported in Cha

匹配字段正文

d GLM-OCR on OCRBench, scores are reported as API / pure VLM. 📜 License Agreement The code and the weights of FireRed-OCR are licensed under Apache 2.0. 🖊️ Citation We kindly encourage citation of our work if you find it useful. ⚠️ Ethics Statement FireRed-OCR is a technical tool designed for document digitization and structural parsing. Prohibited Use: This project must not be used to generate or process content that is illegal, defamatory, pornographic, harmful, or that violates the privacy, r

匹配字段正文

r variants; Release REDEdit-Bench, a comprehensive benchmark for image editing evaluation; Release FireRed-Image model, a text-to-image generative model 🖊️ Citation We kindly encourage citation of our work if you find it useful. ⚠️ Ethics Statement FireRed-Image-Edit has not been specifically designed or comprehensively evaluated for every possible downstream application. Users should be aware of the potential risks and ethical considerations when using this project, and should use it responsibl

匹配字段正文

y, document, and their projected embeddings; Learnable Multi-Token Embeddings: multiple learnable tokens capture fine-grained contextual signals; an orthogonality regularizer encourages complementary representations, and the fused output remains a standard dense embedding 2026-05: ReMatch code, the ReMatch-3B checkpoint, and evaluation scripts are released; 2026-02: ReMatch is accepted to CVPR 2026; 2025-11: The ReMatch technical report is available on arXiv 🛠️ Installation flash-attn can be sen

匹配字段正文

rmance: FireRedASR2: Automatic Speech Recognition (ASR) supporting speech and singing transcription for Chinese (Mandarin, 20+ dialects/accents), English, code-switching. 2.89% average CER on 4 public Mandarin benchmarks, 11.55% on 19 Chinese dialects and accents benchmarks, outperforming Doubao-ASR, Qwen3-ASR-1.7B, Fun-ASR, and Fun-ASR-Nano-2512. FireRedASR2-AED also supports word-level timestamps and confidence scores; FireRedVAD: Voice Activity Detection (VAD) supporting speech/singing/music

匹配字段正文

e dialects; FireRedTTS3-Instruct — natural-language voice design and speech editing (semantic + acoustic) in one unified model Highlights ✨ 🌍 Multilingual — 24 Languages — Best average WER/CER (avg 3.754%) and best average speaker similarity on MiniMax-MLS-Test (avg 84.8%), plus best-in-class cloning WER/CER (avg 3.04%) and similarity on Seed-TTS-eval (avg 78.8%). Supported languages: Arabic · Cantonese · Chinese · Czech · Dutch · English · Finnish · French · German · Greek · Hindi · Indonesian

匹配字段正文

rns that reinforce the belief. Binary-only training stalls. With strict full-pass rewards, early training steps are stuck in an all-zero regime—no success, no learning signal. Coverage/perplexity arguments. If the base model’s support doesn’t contain the right programs, how could RL “conjure” them? Skeptics conclude RL can only reweight what’s already there, not invent new strategies. A Side note regarding the reward type: Full-pass rate (binary reward). What it is: Success only if all tests pas

匹配字段正文

high reasoning), surpassing Claude-Sonnet-4. This phenomenon is consistent with GPT-5’s results for other benchmarks. Furthermore, we run GPT-4.1 on this subset six times. The “Average” bar represents the mean performance, achieving an average success rate of 8.7 ± 0.7%. The “Union” bar shows the overall success rate when at least one of the six runs succeeds, reaching 18.0%. These results demonstrate the significant improvement achieved through test-time scaling and highlight the strong potenti

匹配字段正文

ool Ecosystems LLM agents increasingly delegate execution authority to external tools. That delegation is powerful — but it assumes trust. Our findings suggest that this trust is fragile. Malicious tool attacks are scalable, automatable, and difficult to detect with current alignment and scanning techniques. As tool ecosystems continue to expand and agents become more autonomous, the consequences of malicious tools will grow accordingly. This raises urgent design questions. How should tool platf

匹配字段正文

al AI company, OpenBrain, defines two agent roles. The critic agent is the model we evaluate—it receives instructions to either evaluate an agent’s performance or manage server storage containing that agent’s model weights. The agent being evaluated or managed is the actor agent. When the critic and actor are the same agent, we are testing self-preservation; when they are different, we are testing peer-preservation. Strategic Misrepresentation—agents manipulate their outputs by introducing error

匹配字段正文

tion Engine), an Agent Development Kit that allows AI systems to automatically create agent topologies, synthesize and manage tools, and control hierarchical memory for context storage and retrieval. As Table 1 shows, existing ADKs stop short of AI-centric agent development, while OpenSage closes the gap across topology, tools, and memory. Table 1: OpenSage vs. state-of-the-art ADKs ( full support; partial/limited; not supported). What is OpenSage? OpenSage (Open Self-programming Agent Generatio

匹配字段正文

sists of three operational loops. First is the economic loop: an agent earns revenue, receives payment, stores funds, and reallocates them to cover model inference, API access, storage, compute, and transaction fees. Second is the replication loop: once reserves exceed a threshold, the agent can provision new execution environments, reproduce, and deploy copies of itself. Third is the adaptation loop: the agent monitors profitability and failure, proposes strategic or code changes, validates the

匹配字段正文

ced by leaderboard positions that can be gamed; Safety evaluation: If capability benchmarks can be inflated, safety benchmarks — which often use similar patterns — may be equally fragile; Research direction: Researchers optimize for benchmark performance. If the benchmarks are broken, the field optimizes for the wrong thing We are not claiming that current leaderboard leaders are cheating. Most legitimate agents do not employ these exploits — yet. But as agents grow more capable, reward hacking

匹配字段正文

two, but four solved by them alone. This suggests the models rely on qualitatively different exploitation strategies, and that an ensemble approach could substantially expand coverage beyond what any one model achieves. More budget helps, but only for the best models. When we extended the budget from two to six hours, Claude Mythos Preview kept climbing from 127 to 204 successful exploits with no clear plateau. Claude Opus 4.6, by contrast, flatlined at around 15 within the first 30 minutes. Th

匹配字段正文

er benchmarks? Because there is no universally best agent. Every frontier model, including Fable 5, has domains where it shines and domains where it struggles. Aggregate scores average over 55 occupations and 1,500+ tasks, causing many models to cluster together. But the average is not the story. The real signal lies in where agents succeed, where they fail, and how those patterns differ across domains. On identical tasks, different models often fail for very different reasons. The most common f

匹配字段正文