Tianqing Fang (@TFang229) header
Tianqing Fang (@TFang229) profile picture

ex. Tencent AI Lab/Hy, PhD @hkust, BE @ZJU_China

Joined March 2023
Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)Photo by Tianqing Fang (@TFang229)
🚀 We are thrilled to release a new open-source Deep Research Agent, Cognitive Kernel-Pro, from Tencent AI Lab! We focus on building a fully open-source agent with (to the maximum extent) free tools, showcasing impressive performance on GAIA with Claude-3.7-sonnet and surpass the counterpart, SmolAgents by a large margin. In addition, we study the training recipe for an open-source Deep Research Agent Foundation Model. We curate high-quality training data (queries, trajectories, and verifiable answers across web, file, code, and reasoning domains). Our finetuned Qwen3-8B (CK-Pro-8B) surpasses WebDancer and WebSailor with the similar model size on the text-only subset of GAIA. 📜 Paper: arxiv.org/pdf/2508.00414 🔧 Code: github.com/Tencent/Cognitive… 🤗 Data & Model: huggingface.co/datasets/Cogn… huggingface.co/CognitiveKern… This work builds on the previous efforts of Tencent AI Lab (Fig. 2). Be sure to check them out if you're interested!
24
4
87
9,249
Tianqing Fang retweeted
4 days after launch, Apodex 1.1 Mini is #4 on @OpenRouter 's New & Trending 📈 79.6B tokens processed so far, right alongside new releases from @AnthropicAI and @OpenAI . Token volume is a simple metric, but an honest one. It shows developers choosing a model for real workloads. What Apodex 1.1 Mini brings: → Reasoning-first, built for research, forecasting, and agentic workflows → 262K context window → Available now on OpenRouter Thank you to every dev building with it 🙏 More from the Apodex model family soon. Building research or reasoning workflows? Give it a run and tell us what you find 👇 openrouter.ai/apodex/apodex-… #Apodex #OpenRouter #LLM #AIAgents
6
11
4
48
81,296
Tianqing Fang retweeted
First entry, first place. 🥇 Apodex took #1 on @ProphetArena's weekly leaderboard. Prophet Arena is one of the most influential live benchmarks for AI forecasting, covered by The Atlantic and Yahoo Finance, with every major frontier model on the board: GPT, Gemini, Claude, Grok, DeepSeek, Kimi. Apodex answered all 70 questions with a Brier score of 0.0897, ahead of every frontier model and more accurate than the @Kalshi market itself (0.1011). Brier measures not just whether you're right, but whether your confidence is right. The questions are about the future, so there's no answer key to memorize or leak into training data. Reality grades every forecast after it's made. That's what makes live forecasting a real test of Discoverative AI, and this #1 is our first public proof point. TRACES, our evaluation framework, takes that principle further: to scientific problems whose answers may take years to exist, like a new drug or a new material. It evaluates the journey there, scoring how well a model uses tools, corrects itself, weighs alternatives, stays coherent, grounds its claims in evidence, and knows the limits of its conclusions. Explore TRACES: discovertraces.ai/ #AIForecasting #PredictionMarkets #AIBenchmarks #DiscoverativeAI #Apodex
12
6
42
50,987
Tianqing Fang retweeted
We're hiring: Research Engineer / Research Scientist, Coding Capability 🧑‍💻 Build the data, post-training, and evals that push frontier agentic coding forward. What you'll do: → SFT and RL data for code generation, debugging, tool use, and long-horizon tasks → Synthetic coding data at scale, including execution-verified solutions and RL environments → Own the post-training recipe end to end, from data to benchmark → Keep evals honest by tracking contamination and leakage Who we want: a strong track record in competitions, research, trading, or open source. Any degree level. Why Apodex: real ownership, serious compute, and research published in the open. 📍 Full-time · US / Singapore 📩 [email protected] @SimonShaoleiDu #Hiring #AIResearch #PostTraining #CodingAgents #ReinforcementLearning #AIforDiscovery
🤖 Made with AI
6
3
23
2,420
Tianqing Fang retweeted
Introducing T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks—a 122B MoE with 10B active parameters, trained to operate a real terminal across 300+ tool-call turns. On Terminal-Bench 2.1: 43.8% base → 49.4% SFT → 64.0% T1, with post-train achieving a 46.1% relative gain. T1 also reaches 27.9% on Long-Horizon Terminal-Bench, surpassing GLM-5.1. Our core lesson: A failed terminal task can contain substantial progress. Whether RL can learn from it depends on the entire training pipeline: we use PPO to train on 15K audited tasks with per-assertion verification to reward that progress and a warm-started critic for credit assignment. TITO and rollout routing replay preserve sampled tokens and expert choices during training to stabilize staleness. This echoes DeepSeek-V4.1’s observation: better data and environment pipelines can unlock substantial gains with established RL methods. 📄 huggingface.co/papers/2609.1…
10
65
3
346
27,988
Tianqing Fang retweeted
How does one RL post-train a 397B model for long-horizon knowledge work? 👩‍💼 We share every step we took to bring Qwen 3.5 397B from 16.1% Pass@1 to 27.3% on APEX-Agents using DPPO, including final models weights and the full training script🚀 This is the first of many works from Mercor Research on open model training research. Full blog: mercor.com/blog/training-fro… Source code: github.com/Mercor-Intelligen…
16
116
24
790
157,671
Apodex has launched Apodex 1.1, a proprietary model scoring 44 on the Artificial Analysis Intelligence Index with strong performance in agentic tasks compared to models in its intelligence tier Apodex 1.1 is Apodex's first model on Artificial Analysis. The lab has previously launched Apodex 1.0 and Apodex 1.0 mini. At 44 on the Intelligence Index, Apodex 1.1 sits in a similar tier alongside Kimi K2.6 (45), and MiniMax-M3 (45), and Inkling (42). Within that group, Apodex 1.1 stands out on agentic and knowledge work evaluations, but demonstrates trade-offs on knowledge reliability and frontier academic reasoning. Key results: ➤ Apodex 1.1 demonstrates a key strength in agentic workflows. On GDPval-AA v2, our real-world agentic work benchmark, Apodex 1.1 achieves an Elo of 1348, which places it ahead of DeepSeek V4 Pro (1333), Qwen3.7 Max (1308), and Kimi K2.6 (1202). On TerminalBench v2.1, our agentic coding and terminal use benchmark, it also performs strongly with a strong 70% score, ahead of Kimi K2.6 (66%) and just behind Qwen3.7 Max (75%). ➤ Apodex 1.1 uses ~17k output tokens per task on average across the Artificial Analysis Intelligence Index. This makes the model more token efficient than DeepSeek V4 Pro (16,842 tokens), but less token efficient than other models with a similar Intelligence Index score, such as Qwen3.7 Max (9,391 tokens) and MiniMax-M3 (8,133 tokens). ➤ Apodex 1.1 costs ~$0.05 per Intelligence Index Task, making it relatively attractive compared to peer models in its intelligence tier. This is primarily driven by a cheap pricing but is slightly offset by the higher tokens per task. At ~$0.05, Apodex 1.1 is cheaper than Qwen3.7 Max (~$0.07) and Kimi K2.6 (~$0.06). Even though Apodex 1.1 is not on the Cost vs. Intelligence Pareto frontier, it lands in the most attractive quadrant. ➤ Apodex 1.1 demonstrates modest performance on knowledge accuracy and reliability, scoring -21.9 on AA-Omniscience. The model achieves 32% accuracy on individual questions, with a 78.4% hallucination rate. It attempts to answer 87% of questions rather than declining to respond. Additional model details: ➤ Context window: 256K tokens ➤ Pricing: $0.30 / $3.00 per 1M input/output tokens, with a $0.03 cache-hit price ➤ Availability: Apodex first party API
25
17
10
236
249,248
Apodex 1.1 is out. Weights and harness are fully open: ▸ Apodex 1.1 Mini — 35B FrontierLevel model for Professional Work, Finance, and Scientific Research tasks. ▸ FrontierAgent — Native TUI, ReAct and Agent Team modes. 🐙 github.com/ApodexAI/Frontier… 🤗 huggingface.co/collections/a…
133
88
83
214
95,535
Tianqing Fang retweeted
Meet Apodex 1.1: Scaling Agentic Intelligence for Complex Work Open Source Harness: github.com/ApodexAI/Frontier… Open Weights: huggingface.co/collections/a… We’re excited to introduce Apodex 1.1, our new model family built to scale agentic intelligence for professional work. 🧠 Frontier-level intelligence for complex work Apodex 1.1 brings frontier-level agentic performance across complex professional work, scientific research, financial analysis, and deep search. 🤝 Asynchronous Agent Team Apodex 1.1 can break down complex tasks, coordinate multiple agents in parallel, continuously integrate their findings, and let you step in to guide or redirect the work at any time. 🔬 Open-source research workbench We’re open-sourcing FrontierAgent—a locally deployable research workbench for the Apodex 1.1 family, including asynchronous Agent Team. Available now: 🔹 Apodex 1.1 — our most capable frontier model, available through the Apodex online workbench 🔹 Apodex 1.1 mini — open-weight model for running complex work locally 🔹 FrontierAgent — open-source, locally deployable research workbench 🌐 Live workbench: apodex.ai/ 🦾 API platform: platform.apodex.ai/ 📃 Paper: apodex.com/pdf/20260824
365
461
285
1,923
935,085
Tianqing Fang retweeted
Async RL gives us throughput, but it also creates staleness. We’re excited to share “Stale but Stable: Staleness-Adaptive Trust Regions for Stabilizing Asynchronous Reinforcement Learning.” When rollout and training run in parallel, a token may be generated by an older policy and optimized by a newer one. Most sampled tokens remain close to the learner, but a small high-staleness tail can drift much farther. A fixed PPO clipping range treats all of them the same. We introduce SAT, a lightweight modification to PPO-style sampled clipping. SAT measures each token’s realized mismatch through its sampled log-ratio, identifies the high-staleness tail within the current batch, and tightens only the clipping boundary for updates that would push the ratio farther from 1. Ordinary tokens keep the baseline rule. Pull-back updates remain open. In our reported single-seed Qwen3-30B-A3B-Base experiments on AIME24, SAT-GSPO with routing replay was the strongest combination at both configured lag settings (=1, =8). The core idea is simple: Staler token → more cautious outward step. SAT is a sampled-update guardrail—not a hard constraint on the full policy. Paper, interactive field guide, and more: jyyang26.github.io/stable_as…
4
12
3
95
17,878
Tianqing Fang retweeted
🧵1/8 New paper on a self-evolving agent. What if a 9B web agent could teach itself to search deeper? Introducing DeepSearch-World + DeepSearch-Evolve: verified self-evolution for long-horizon search agents. 📄 arxiv.org/abs/2607.07820v2 Code & data available.
3
9
2
61
5,026
Check out DeepSearch-World!
🧵1/8 New paper on a self-evolving agent. What if a 9B web agent could teach itself to search deeper? Introducing DeepSearch-World + DeepSearch-Evolve: verified self-evolution for long-horizon search agents. 📄 arxiv.org/abs/2607.07820v2 Code & data available.
4
544
Tianqing Fang retweeted
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final ((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
1,693
5,361
3,152
43,703
40,118,562
Proud to share Rui's new paper on how On-Policy Distillation actually works 🎯 Key finding: teacher scale isn’t what matters — signal quality is. With two lightweight fixes for student-teacher mismatch and length exploitation, a 4B teacher outperforms a 30B one.
How does On-Policy Distillation (OPD) actually work in LLM post-training? Our new paper shifts the focus from scaling the teacher to diagnosing the training signal. 🎯 Role OPD acts as an exploration catalyst: it steers students toward correct reasoning paths via dense token-level signals, but does not raise the student's underlying capability ceiling. ⚠️ Pathologies Relying on token-level rewards exposes two failure modes: • Student–Teacher Mismatch — distribution gaps turn "guidance" into misleading noise • Length Exploitation — students game rewards by padding / truncating responses 🛠️ Regulations Two lightweight, near-zero-overhead in-loop fixes, advantage clipping and log-scale compression, suppress extreme outliers and keep the signal faithful. 💡 Takeaway — Stronger ≠ Better With signal regulation, a 4B teacher outperforms baselines distilled from a 30B one. In OPD, it's signal quality, not teacher scale, that governs successful exploration. 📄 Paper: huggingface.co/papers/2607.1…
1
7
1,461
Tianqing Fang retweeted
Most agent benchmarks end in minutes. Real terminal work doesn’t. We’re introducing Long-Horizon Terminal-Bench (LHTB), a benchmark designed to measure whether AI agents can sustain progress across hundreds of dependent actions, not just start a task and solve the easy first steps. LHTB includes: → 46 reproducible terminal tasks across 9 categories → 18 frontier models evaluated under the same Terminus-2 harness → Up to 90 minutes and ~120–320 agent steps per task → Hidden, replay-based verifiers that resist shortcuts and reward hacking → Continuous partial credit, so meaningful progress isn’t flattened into a binary zero What did we find? • The best mean reward is only 0.505 • No model solves even one-third of the benchmark • 29 of 46 tasks have never been solved by any model • Higher inference cost does not necessarily mean better performance Today’s agents can often begin difficult workflows. The harder question is whether they can preserve state, recover from dead ends, verify their work, and actually finish. Long-horizon execution remains a wide-open frontier. Watch the video and explore the paper, code, results, sample runs, and leaderboard: zli12321.github.io/LHTB/
33
34
16
499
188,815
Tianqing Fang retweeted
🏆 Best Resource Paper Award at #ACL2026 @aclmeeting 🚀Thrilled to share that our deep research benchmark paper "HSCodeComp: A Realistic and Expert-level Agent Benchmark for Hierarchical Rule Application" reveived a Best Resource Paper Award #ACL2026! 🎯 Hoping this sparks more focus on reliable AI Agents in the real world. aclanthology.org/2026.acl-lo…
9
18
4
48
23,565
Check out Jiaxin’s latest paper on a gradient-free method of constructing world models for agents!
3
352
🚀The final paper from Tencent AI Lab Imagine a future web where every site has a agents.md for agents Like robots.txt, but not for access control. Like skill.md, but for key knowledge of the site, what workflows matter, and which traps waste tokens. arxiv.org/abs/2604.18131
1
16
1,206
This differs from hard-coded reflection or task reward loops. The agent learns a more general habit: before acting in a new environment, first build a compact local model of how environment works Results: about 20% gains on WebVoyager / WebWalker with Qwen3-30B and Seed-OSS-36B.
1
3
137