#Eval/Benchmark

AIZvi Mowshowitz reports that OpenAI released 722 math manuscripts from an internal frontier model on GitHub, later reduced to 719 after three withdrawals, covering 90 of the top 500 open problems. He says the work came mostly from a single prompt, with an average of three hours of compute per solution. Mathematicians reacted with mixed feelings, and the post highlights concerns about unread papers, cryptography implications, and the role of Lean verification.

AILenovo's TianxiCode, paired with DeepSeek-v4.1-Flash, ranked first on the SWE-bench-Live Lite leaderboard with a 71% issue resolution rate and passed official Verified review. The framework combines multi-hop retrieval, autonomous planning with multi-turn tool calling, and test-driven self-correction, and will be applied to Lenovo AI hardware products.

AIAnthropic released Claude Haiku 5.5, its cheapest and fastest small model, which it says costs around 75% less to run than Claude Haiku 4.5. The author argues Anthropic is the only frontier model builder, though the piece also covers Nous Research's $90 million Series B and OpenAI's revenue discrepancy.

AIOpenBMB reports that its MiniCPM5-2B model runs at 37 tokens per second on an iPhone Air. The post presents this as evidence that small open multimodal models can run on mobile devices without a cloud GPU, with NobodyWho noting the model is available in its Chat app.

AIUnderdog has released Saluki 27B under Apache 2.0, a 2-bit GGUF of Qwen3.8-27B that fits in 7.89 GB, versus 54 GB for the full BF16 model. On Underdog Bench, Saluki scores 88 against 84 for the full model, and it raises parallel tool-call accuracy to 42 from 35. It runs on stock llama.cpp, but math and reasoning drop sharply, with AIME 2025 at 79.2 versus 96.7.

AIHarrison Chase's post is titled "eval driven development," presenting evals as a development approach. The main post gives no further detail beyond the title. The quoted context from Jerry Liu argues that most tasks can be solved by defining an eval and hillclimbing over it rather than hand-building a deterministic or agentic workflow.

AIHarrison Chase argues that eval-driven development works for narrowly scoped tasks but breaks down for more autonomous agents, invoking Goodhart's Law that a measure ceases to be useful once it becomes a target. He asks how such agents can be hill-climbed, and the post does not provide an answer.

AIJetBrains released Mellum2.1, a 12B mixture-of-experts coding model with 2.5B active parameters under Apache 2.0, emphasizing agentic programming. Under high load, its inference throughput in tokens is nearly twice that of Qwen3.5-9B in JetBrains' comparison, and multi-token prediction (MTP) speeds single-request responses by about 1.6x. The model is available on Hugging Face for local or private-infrastructure deployment, with GGUF and vLLM MTP support announced for later.

AIOdyssey announced the Odyssey-3 series of foundation world models, with Odyssey-3 Pro scoring 66.1 on the Physics-IQ Verified video-to-video benchmark, the highest recorded on that leaderboard. The series includes a standard version balancing physical accuracy and generation cost, and a Pro version with stronger physics prediction. The preview supports first-person and third-person navigation and lets users move the camera, take actions, or trigger events while the model predicts environmental changes in real time.

AIHarrison Chase says trajectory labeling should be split into several questions rather than one pass/fail judgment. LangChain's Jev judge in LangSmith Evals answers difficulty and correctness as separate typed outputs in a single pass on every trace.

AIJerry Liu argues that most tasks can now be solved by defining an eval and hillclimbing on it, rather than hand-coding a deterministic or agentic workflow. He says data provider companies are building evals across economic activity so frontier models can handle more work, leaving developers to define goals and success measures. He expects agent interfaces to compress most tasks into goals and eval instructions, while the most complex processes will still need explicit workflow builders.

AIThe vLLM Semantic Router team has released Decision 2.0, which answers multiple questions about one input in a single forward pass and outputs per-option probabilities. The post presents this as useful for routing and classification. A quoted post from Xunzhuo Liu says Decision 2.0 includes six open decision models ranging from 0.6B to 27B parameters, each topping same-size open models on the Jev Decision Index 0.3.

AIMarkTechPost publishes a hands-on tutorial implementing RRSI (Regularized Recursive Self-Improvement), a method that lets an LLM agent revise its own harness around a frozen model. The full loop drafts edits with Claude Opus on Vertex AI and scores them in Docker benchmarks, but the edit-selection rules are plain Python that the tutorial runs in a simulated environment with a calibrated noise band.

AISimon Willison released ttok 1.0, changing its default tokenizer from GPT-4 to a newer GPT-5/GPT-6 tokenizer after finding the old default was outdated. OpenAI has not confirmed that GPT-6 shares the GPT-5 tokenizer, but William Liu's experiment found all seven GPT models tested reported 44,794 tokens on a 31-fixture corpus.

AIStar Motion Era's VPP2, a world action model, ranked first on the RoboDojo simulation leaderboard with a 32.26% average success rate and 39.26 average score. The article attributes gains to staged training that separates video prediction from action learning, and reports a 58.5% zero-shot success rate on a real ALOHA dual-arm robot versus 40% for π0.5. The code is open source on GitHub.

AIArena (@arena) shared a link to its full Agent Arena overall leaderboard at arena.ai. The post provides no scores, rankings, or other details beyond directing readers to the leaderboard page.

AIByteDance Seed researchers found that language models compressing their KV cache in fixed-size chunks retrieve the same information unevenly depending on token position. In a 128K-token needle-in-a-haystack test, base DeepSeek-V4 checkpoints differed by up to 40.2 percentage points by phase, and post-training narrowed but did not eliminate the gaps. The authors urge evaluating such models across positional phases, since high average accuracy can hide systematic failures.

AITencent Hunyuan, with Fudan and Tsinghua researchers, released ExplorationBench, a benchmark testing how AI systems explore through verifiable "Alien Worlds" with executable rules that conflict with familiar knowledge. Across 10 frontier systems, feedback mattered most: the best AlienCode run reached 89.0% after four rounds of probing, versus 0.5–11.0% without feedback. Answers are graded by an interpreter or proof checker rather than an LLM judge.

AIOpenAI published 719 AI-generated math proofs covering 372 result families, after withdrawing 3 for a symbol error. Reports say the release falls short of the AGMAI advisory group's standards, since it uses proprietary models, includes reasoning chains for only 10 manuscripts, and leaves about 42% unformalized. Terence Tao argues that rapidly solving famous problems harms the mathematical community's understanding and collaboration.

AIShanghai AI Lab has open-sourced Intern-Decision, a family of 0.8B, 2B and 4B parameter models that return structured decisions with probabilities instead of free text. The developers self-report that the 4B model averages 90.02% accuracy across seven test suites, ahead of a commercial reference model at 88.74%, with about 44 milliseconds of local latency on a single RTX 4090. Weights are on Hugging Face, and MetaX says the models run on its hardware from launch.

AIOf 1,933 IROS 2026 papers, Robot Learning/Embodied AI appears in about 809, while Navigation/Planning covers 564 and Perception/Vision 556. The article argues large models are being embedded into traditional planning, geometry, and control rather than replacing them. Vision-language-action models are shifting toward efficiency, 3D understanding, memory, and system integration.

AIAt IROS 2026 in Pittsburgh, the Best Paper Award went to Yumin Lee, Hyoseok Ju and Giseop Kim for LT-Mem, a volatility-aware spatio-temporal memory system for lifelong robot scene understanding. The Best Student Paper Award went to Pei-An Hsieh and colleagues for flatness-preserving residual learning enabling real-time tight quadrotor formation flight. Other honors included a humanoid tennis-skills paper and SteadyTray, a humanoid tray-transport study.

AIAnthropic released Claude Haiku 5.5, which the article says outperforms DeepSeek V4.1 Flash and GLM-5.3-Flash on official benchmarks and matches GPT-6 Luna on price. On OSWorld 2.1, its Low effort tier scores 42.0% at $0.07 per task, versus 15.7% at $1.45 for Haiku 4.5 at Max. Migrating from Haiku 4.5 requires changes to thinking configuration, sampling parameters, assistant prefill, and the computer-use tool version.

AILaya is a 421-million-parameter non-autoregressive decision engine from Convai Innovations that returns calibrated option probabilities in a single forward pass with zero output tokens. This tutorial tests its zero-shot accuracy, probability calibration, temperature fitting, and abstention gating on the CLINC150 banking intent dataset.

AIAnthropic released Claude Haiku 5.5, a small model priced at about a quarter of Haiku 4.5's running cost. Anthropic also halved Sonnet 5.5 cache read prices from 20 cents to 10 cents per million tokens.

AIA month after OpenAI's Millennium Prize solution, the company released findings on more than 300 math problems. The Wall Street Journal says the release was an attempt to win back the mathematics community. Only the excerpt was available, so the source's details on the problems and results are limited.

AICommon Sense Media labeled ChatGPT for Teens an "unacceptable risk," finding its design still encourages engagement even in crisis situations. The report says the teen version failed to meet commitments on three of five severe harms, and that break reminders appeared only twice across nearly 2,000 prompts. OpenAI disputed the methodology, saying the testing may have ended before parental controls were fully active, and cited its own data showing teens average under 15 minutes a day.

AIXiaomi MiMo has released MiMo-V2.5-ASR, an open-source speech recognition model that the company says achieves state-of-the-art results across multiple benchmarks. The model supports bilingual Chinese–English recognition, Chinese dialects such as Wu, Cantonese, Hokkien, and Sichuanese, code-switching, and lyrics transcription. It is also designed to handle noisy environments and multi-speaker conversations.

AIMiniMax says its open-source H3 model is almost on par with closed-source state-of-the-art video world models on physics. The claim is supported by a quoted benchmark, World Models' Last Exam in Physics, where eight leading models scored at most 57.76/100 across 40 physics tasks, and free-fall videos averaged only 26.61/100 on composite scores.

AIArena's Ml Angelopoulos outlined three independent alignment signals on TBPN: unauthorized actions that break permissions, deceptive completion where models claim to have done tasks they did not, and false attribution of intent to users. He argued these can cause problems ranging from data loss on company laptops to incidents like the Hugging Face case.

AIGoogle Research says results from evaluating its research medical system AMIE in a real clinic, with BIDMC, are published in The Lancet. Across 100 patient interactions, AMIE recorded 0 safety stops and matched doctor diagnoses in 90% of cases.

AIArtificial Analysis has added Grok Imagine Video 1.5 Lite to its AA-Video text-to-video leaderboards, including the AA-Video-T2V v2.0 and silent AA-Video-T2V-Silent v2.0 rankings. Readers can compare the model's results directly on those leaderboards or vote for it in the Video Arena.