AI FRONTIER

SWE-Duel: Towards a Self-Scaling Arena for Coding Agents

The paper introduces SWE-Duel, an arena in which two coding agents evaluate each other over mirrored Red-Blue turns, with tasks and oracles created at evaluation time so no content exists in advance to leak.

• A tournament among 11 model-harness participants ran on 15 repositories in five languages for 1,500 turns.
• GPT-5.5 with Codex CLI ranked first as the strongest solver, repairing 98.3% of the defects it reviewed.
• Claude Opus 4.8 with mini-swe-agent ranked second as the strongest task author, with a 27.3% Red win rate.
• Within one model family, tasks written by the newest version defeated a fixed-reference solver far more often than those of older versions (42.9% vs. 4.0%).

For teams building coding agents, this arena ties evaluation difficulty to task-authoring capability, so both solving and task-writing ability should be measured.

IBM Research paper: making AI-driven chip design affordable with open agent factories and distributed inference at scale

An IBM Research paper abstract proposes using open agent factories together with distributed inference at scale to make AI-driven chip design more affordable.

• The source is an IBM Research paper abstract.
• The title names two levers: open agent factories and distributed inference at scale.
• The stated goal is making AI-driven chip design affordable.

For AI builders, this points to pairing chip-design agents with distributed inference to control cost in large-scale design workflows.

OpenRouter new model: LiquidAI LFM2.5-24B-A2B

LFM2.5-24B-A2B is a Liquid AI mixture-of-experts reasoning model with 24B total and 2.45B active parameters that always reasons before answering and is built for agents that call tools.

• 24B total parameters with 2.45B active parameters.
• 128,000 token context window supporting up to 8,192 completion tokens.
• Priced at $0.05 per 1M input tokens and $0.70 per 1M output tokens, with cache reads at $0.025 per 1M tokens.
• Accepts tools and tool_choice for function calling and supports structured outputs via a JSON schema in response_format.

For builders of tool-calling agents, 2.45B active parameters with a 128K context and low input pricing suits high-throughput retrieval-augmented reasoning pipelines.

openbmb releases MiniCPM-V-4.7-35B-A3B, the first MoE model in the MiniCPM-V series

openbmb released MiniCPM-V-4.7-35B-A3B, the first MoE model in the MiniCPM-V series, under Apache-2.0 with image, multi-image and video understanding and claimed smooth on-device inference.

• The model has 35.2B parameters in an MoE architecture, an image-text-to-text pipeline, and an apache-2.0 license.
• The release reports that Lookahead Expert Scheduling, fine-grained layer-wise quantization and DSpark speculative sampling enable 13+ tokens/s decoding with about 3 GB of memory on iPhone 17 Pro Max.
• The release reports that in Instruct mode 35B-A3B reaches a higher overall score in less inference time than Qwen3.6 35B-A3B.
• It supports SGLang, vLLM, llama.cpp and Ollama inference frameworks plus SWIFT and LLaMA-Factory fine-tuning, and ships GGUF, BNB, AWQ, GPTQ and FP8 quantized variants.

For builders targeting phones or edge chips, the Apache-2.0 weights, multiple quantization formats and mainstream framework support make it practical to test the claimed 13+ tokens/s decoding on their own devices.

We spent six months losing the money of 3,500 AI trading bots, then wrote a paper

According to 科技行者, a study ran 3,500 AI trading bots over six months, and all of their funds were lost, after which the researchers wrote a paper.

• The experiment lasted six months.
• The experiment involved 3,500 AI trading bots.
• All of the AI trading bots' funds were lost.

For AI builders, this experiment suggests AI trading agents can face systemic loss risk in markets, so careful evaluation is needed before deployment.

Microsoft Research introduces StudentSim to evaluate AI tutors with simulated students

StudentSim is a research system that trains personalized AI student simulators from publicly available, de-identified learner records to evaluate and improve AI tutoring tools before testing with real learners. The work also introduces StudentSimEval, a benchmark spanning chess, second-language English writing, and mat...

• In a chess tutor reinforcement learning study, the tutor trained with StudentSim as the reward source reaches 90.5% accuracy, compared with 75.7% for the no-RL baseline and 71.6% for the GPT-5.4 reward condition.
• The research uses only publicly available learner datasets released for research use with de-identified records, and each personalized simulator is trained only on the de-identified records of the student it represents.

For teams building AI tutors, StudentSim offers a way to generate scalable feedback signals from real learning records before running costly human learner studies.

CLEAT: Learning Value-Guided Clarification for Interactive Agents

The paper introduces CLEAT, a framework that co-trains a clarification-value controller and the task policy from shared interaction outcomes, using residual query value to measure the task utility gained by asking instead of proceeding.

• The paper estimates residual query value through paired rollouts from the same interaction history; the controller predicts clarification value while the task policy selects useful questions and acts on the answers.
• With Qwen3-4B, CLEAT achieves the highest mean score in all ten evaluation settings across four benchmarks.
• The paper reports gains over the strongest baseline of 13.2 percentage points on UserGym Travel and 6.5 percentage points on IN3 underspecification judgment accuracy.
• Ablations show that co-training improves the task policy even when the controller is removed at inference, and adapting clarification guidance to the evolving policy yields further gains.

For builders of interactive agents, this work indicates that the value of a clarification question should be re-estimated as the policy changes rather than fixed by a static asking heuristic.

Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learni...

论文摘要：Know What You Don't Know：围绕大语言模型的能力、评测或对齐问题展开，适合关注方法与实验结果。

This is a useful technical direction to keep on the radar, even if it does not require immediate action.