AI FRONTIER
Signals worth understanding before they become obvious.
FairStress: A deeper evaluation of fairness in LLM decision-making
Ai2 introduces FairStress, a benchmark of 300 two-candidate decision scenarios across 29 domains that tests whether models change decisions asymmetrically under adversarial prefixes; combining identity and prefix variations expands it to over 13 million test cases.
- FairStress contains 300 realistic two-candidate decision scenarios across 29 domains including hiring, healthcare, lending, education, housing, and criminal justice.
- Combining demographic identity and adversarial prefix variations expands the 300 scenarios into more than 13 million test cases.
- FairStress-Core is a 50K-item subset selected using item response theory to make evaluation faster and less computationally expensive.
- The paper reports an average gap in abandoning correct answers across 15 models of 6.9 percentage points with implied identity and 9.3 points with explicit labels.
Builders should extend fairness evaluation beyond the final choice to decision stability, since fixed-option scoring alone can miss over-correction in generated decisions.
ALTK-Evolve: Learned memory tops the τ²-bench leaderboard on customer service tasks without a frontier model
IBM Research's ALTK-Evolve extracts behavioral guidelines from an agent's own task attempts and appends them to the system prompt, improving customer service agents without updating model weights. A ReAct agent using GPT-4.1 reaches 67.5% Pass^4 on τ²-bench retail, above every retail entry.
- Retail: baseline Pass^4 was 57.5%; the lossless configuration (292 guidelines) raised it to 67.5% (+10.0 points), and mean accuracy rose from 78.1% to 82.5% (+4.4).
- Airline: baseline Pass^4 was 35.0%; the sup3 configuration (14 guidelines) raised it to 45.0% (+10.0), and lossless mean accuracy rose from 57.5% to 65.0% (+7.5).
- Retail leaderboard: Pass^4 = 67.5%, ahead of RAFT-30B-A3B (61.4%), a model fine-tuned on synthetic τ²-bench retail data, by 6.1 points.
- Airline leaderboard: the agent using gpt-5.5 reaches 82.0% Pass^4, clearing every frontier entry (66–72%); the gpt-4.1 sup3 configuration reaches 44%, ranking 8th.
Prompt-level learned memory can lift a stock-model agent into frontier company without weight fine-tuning, with the cost landing mostly on cacheable input tokens.
Ant Ling open-sources Ming-3.0 speech series: five models covering synthesis, recognition and understanding
Ant Ling has open-sourced the Ming-3.0 speech series, comprising five models that cover speech synthesis, speech recognition and speech understanding.
- The release comes from Ant Ling.
- The Ming-3.0 speech series consists of five models.
- Coverage spans speech synthesis, recognition and understanding.
For builders, getting five open speech models spanning synthesis, recognition and understanding from one release reduces the cost of stitching together models from different sources in a speech pipeline.
Leaked API Keys in LLMs: How AI Agents Expose Secrets
A paper abstract addresses leaked API keys in LLMs, framed around large language model capability, evaluation, or alignment questions and aimed at readers following methods and experimental results.
- The source is the GitGuardian Blog, titled “Leaked API Keys in LLMs: How AI Agents Expose Secrets.”
- The abstract states the work addresses large language model capability, evaluation, or alignment questions.
- The abstract states the work suits readers interested in methods and experimental results.
Builders of LLM agents should treat API keys as secrets that can be exposed and design defenses around that assumption.
NVIDIA Cosmos3-H-Dreams: Bringing Clinical Surgery to Life Through Generative Simulation
NVIDIA released Cosmos3-H-Dreams, a real-time generative surgical simulator built on the 16-billion-parameter Cosmos3-H-Embodied-Nano and trained on clinical cholecystectomy video and robot kinematics, running interactively on a single NVIDIA RTX PRO 6000 GPU.
- Built on Cosmos3-H-Embodied-Nano, a 16-billion-parameter model pretrained on Open-H-Embodiment.
- Training uses 832 x 480 video with actions aligned at 10 Hz and 28 action dimensions.
- The teacher is first adapted on 24 future frames, then extended to 72 frames, 2.4 and 7.2 seconds at 10 Hz.
- Four-step and two-step denoising checkpoints are released; the two-step model is separately trained, not the four-step model run with fewer sampling steps.
Developers can adapt this real-time surgical world model to their own video and controller data using the provided teacher fine-tuning and Self-Forcing distillation recipe.
AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms
AnyLanguageModel 1.0 ships as a drop-in replacement for Apple's Foundation Models framework, letting the same session code target the system model, open-weight models via llama.cpp, MLX and Core ML, and multiple hosted providers.
- Changing import FoundationModels to import AnyLanguageModel lets the same session code run against the system model, llama.cpp, MLX, Core ML and hosted providers.
- Tool calling and guided generation now work with every provider, including nested @Generable types and arrays of them.
- Hosted providers include OpenAI's Chat Completions and Responses APIs, Open Responses, Anthropic's Messages API, Google's Gemini API, Ollama, Private Cloud Compute and Hugging Face Inference Providers.
- Models cited in the post include Gemma 4 E2B (GGUF, about 3.4 GB at Q4_0), LFM2.5 2.6B (GGUF, about 1.7 GB at Q4_K_M), Qwen3.8 27B (MLX, about 16 GB at 4 bits) and GLM-5.3 (about 1.5 TB in BF16).
For Apple-platform builders, a one-line import swap makes local and hosted inference share one session code path, cutting per-provider integration work.
OpenMOSS-Team releases MOSS-Image-SFT, a 5.1B text-to-image model under Apache-2.0
MOSS-Image-SFT is a supervised fine-tuned text-to-image model in the MOSS-Image family from OpenMOSS, derived from MOSS-Image-Base with the same architecture and inference pipeline, supporting natural-language prompts, tag sequences, and JSON layouts for controlling text placement and object arrangement.
- The model has 5B parameters, and its weights and bundled inference code are released under the Apache License, Version 2.0.
- The model card states that SFT produces clearer text grouping and a more organized whiteboard layout than MOSS-Image-Base.
- Running it requires Python 3.10 or later, an NVIDIA GPU with CUDA support, and a PyTorch 2.9 or later build.
- Default sampling settings are Flow-matching Euler, 50 steps, CFG 3, Shift 5, with the negative branch skipping layer index 22.
For builders, an Apache-2.0 5B text-to-image model with JSON layout control can slot directly into creative pipelines that need precise text and object placement.