EvoCode-Bench evaluates whether coding agents can adapt their own codebases across rounds as requirements change, while meeting new and still-active prior requirements. GitHub：github.com/UniPat-AI/EvoCode… arXiv：arxiv.org/abs/2605.24110

Professional work spans apps and depends on earlier actions. SaaS-Bench evaluates computer-use agents on 106 realistic tasks across 23 deployable SaaS systems in 6 professional domains. Experimental results show that even the best-performing model achieves an end-to-end task completion rate of less than 4%, revealing limitations in agents’ abilities to plan tasks, track states, maintain cross-application context, and recover from errors. GitHub：github.com/UniPat-AI/SaaS-Be… arXiv：arxiv.org/abs/2605.15777

RoadmapBench evaluates coding agents on 115 long-horizon, multi-target tasks based on real open-source version upgrades across 17 repositories and 5 programming languages. The results show that even the best-performing model, Claude-Opus-4.7, completed only 39.1% of the tasks, while the lowest-performing model achieved a completion rate of just 5.2%. GitHub：github.com/UniPat-AI/Roadmap… arXiv：arxiv.org/abs/2605.15846

Three papers from UniPat AI and our collaborators have been accepted to #NeurIPS2026. RoadmapBench · SaaS-Bench · EvoCode-Bench Building the Experience Layer between AI and the real world. Explore the work below. See you at NeurIPS!

Validation has the lowest mean stage score in PaperBenchX. Mean stage scores across all 10 configurations: – Modeling: 62.73% – Execution: 61.83% – Validation: 42.91% Running an experiment is only part of reproducing its scientific result.

A plausible result is not enough. We remove generated outputs and rerun the submitted workflow in a clean, offline environment. Only regenerated artifacts can support rubric credit. Scoring combines programmatic checks with a unified scientific grading model.

Before a paper enters the evaluation, its reproduction scope must be clearly defined, a domain-specific environment configured, and the reference workflow and scoring rules validated by experts. PaperBenchX divides this process into two stages—“Candidate Task Construction” and “Validation and Review”—comprising nine steps in total. Through this process, the scientific requirements of each paper are translated into clearly defined tasks, runnable reference workflows, and scoring criteria that can be assessed item by item.

PaperBenchX comprises 93 reproduction tasks derived from 93 research papers, spanning 12 research domains and 10 domain-native scientific environments. It includes 3,168 expert-validated scoring criteria, with PhD-level domain experts contributing over 1,000 hours in total to the construction and validation of the benchmark.

On 93 tasks, PaperBenchX finds that GPT-6-Astra (Codex) leads 10 model-harness configurations: – Mean rubric score: 59.08% – Full reproduction rate: 13.98% (13/93) Full reproduction requires a task score >90%. Substantial partial progress rarely becomes complete reproduction.

PaperBenchX selects one representative task from each research domain and open-sources 12 test tasks, allowing the community to run evaluations, examine the evaluation methodology, and investigate the capabilities and limitations of scientific research Agents. It also maintains 81 closed test tasks with limited access to prevent evaluation contamination. Run your agent. Inspect the evidence.

Introducing PaperBenchX.🥳 A benchmark for measuring Agents’ ability to reproduce scientific research end to end. 93 tasks. 12 research areas. 10 domain-native environments. 3168 expert-validated criteria. First step toward General AI Scientist GitHub: github.com/UniPat-AI/PaperBe… Blog: unipat.ai/blog/PaperBenchX

Echo is live. Our prediction intelligence system is now running in production, turning uncertainty into measurable outcomes. Prediction should be general, evaluable, trainable, and profitable. Echo is how we get there. Developer API coming soon. Stay tuned.

[9/10] Echo outperforms the human market: 🧠⚔️📊 🏛️ 63.2% in Politics & Governance 📅 59.3% on 7+ day horizons 🌫️ 57.9% when the market is uncertain

[8/10] EchoZ API delivers calibrated probabilities, evidence, counterfactual analysis, and monitoring recommendations. 📡 Built to capture alpha. In the last two weeks, 4 of 5 OpenClaw bots using our API profited on Polymarket. 📈 Join the waitlist: echo.unipat.ai/apply

[7/10] The lead is robust. 📈🛡️ Across the full σ sensitivity sweep, EchoZ stays #1. The benchmark is also designed to remain stable under: 🔄 missing submissions 🧊 cold starts 🌊 changing model pools

[6/10] On the March 2026 Echo leaderboard, EchoZ-1.0 ranks #1 with 1034.2 Elo — ahead of Gemini-3.1-Pro, Claude-Opus-4.6, Grok-4.1-Fast, and GPT-5.2.

[5/10] Trainable At the core is EchoZ-1.0 — the first LLM trained end-to-end under the Train-on-Future paradigm. 🚀 The core mechanisms include: 🧪 Dynamic Question Synthesis 🔍 Rubric Search 🗺️ MapReduce Agent Architecture

[4/10] We rethought prediction evaluation. 📊 Prediction gets easier as new information arrives, so comparing models at different timestamps is noisy. Echo evaluates models in pairwise battles, aligned on the same question at the same prediction time. 🎯⏱️

[3/10] Echo has 3 layers🧩: — a dynamic evaluation engine — a Train-on-Future post-training paradigm — an AI-native prediction API