A 9B small language model running on one local GPU performs case-level and trio-based variant interpretation, recovering 93% of expert-reported causal variants across 327 solved singleton visual-impairment cases and classifying 82% of them as pathogenic or likely pathogenic.
ScienceRSI splits scientific capability into specialization, verification, transfer and reuse, so an expert model contributes feedback only where it is genuinely better on the same scientific requirement. In the 4B-to-4B setting, scientific-reasoning average reached 55.0%, 2.0 points above GRPO; transferring from Qwen3-8B experts to a Qwen3.5-27B target raised the equal-weight average over seven benchmarks from 87.3% to 94.0%. Removing expert-contribution verification dropped the same 4B student's scientific-reasoning average from 55.0% to 50.6%.
WebAnything automatically rebuilds real URLs into executable, resettable local website environments without per-site manual construction. On five WebHarbor-matched sites and 14 historical states it scores higher on pixel similarity, SSIM and Text F1, with a 12-2-0 blind-audit record; fine-tuning Qwen3.5-9B on 5,856 of its interaction trajectories raises REAL success by 9.9 percentage points.
Base Labs released joey-1, a 2B decision model that answers label-style calls such as routing a ticket or flagging a message by returning a single token and a calibrated probability in one forward pass, instead of generating text. At production scale it serves the average 2,000-token request at a P50 end-to-end latency of 20 ms, about 2.5x faster than elephant-1 at 49 ms, and at $0.01 per million tokens it costs a tenth of mammoth-1.
AI research agents must train and evaluate a model for every candidate program, so execution cost caps how many candidates they can compare per step. MLE-World, finetuned from Qwen3.5-4B, reads a program's code and data description and predicts its terminal output, exception type and validation score, plus a confidence score. On held-out programs from 17 MLE-bench Lite tasks its predicted terminal output matches the real run on 77.9% of programs versus 41.1% for the strongest prompted baseline, and search guided by it reaches up to 22.1% higher observed mean performance than the agent without a world model while using less real execution.
Baseten's research lab released two open-source decision models, mammoth-1 (27B) and elephant-1 (9B), which return a probability distribution over every possible answer instead of generating text. mammoth-1 ranks first among 208 decision models on the Jev Decision Index 0.4, while elephant-1 is tied for eighth there and is the leading 9B decision model, and is tied for first with d1-ultra on the separate Artificial Analysis Decision Index. Both were post-trained from Qwen3.5-9B and Qwen3.5-27B on 6.21 million decision-making examples, are free to try for a week on Baseten Model APIs, and are open-source on Hugging Face.
Theory-of-mind benchmarks usually ask story questions about what others believe, and a correct answer does not show the model can use that judgment in a task. This evaluation extends household collaboration and calendar scheduling with asymmetric information and separates world understanding, information access, beliefs, decisions and execution into independent interfaces. On matched scheduling cases Claude recovers all agent beliefs yet produces subjectively optimal plans in only 67.9% of cases, while supplying beliefs raises Qwen-27B household task success from 55.0% to 75.0%.
Physical hardware security evaluation has depended on a person choosing each laser fault-injection position, power and timing. A large-language-model agent now links firmware reverse engineering, chip imagery and experimental decisions into a closed loop, choosing the next shot inside human-approved safety bounds. On an ATmega328P it produced a skipped load instruction and a wrong-PIN acceptance confirmed by a separate state read; on an Artix-7 FPGA the same workflow extended to ciphertext-error classification in programmable logic.
In a Mesa implementation of the Schelling segregation model, one agent's symbolic rule for classifying neighbors is replaced by a locally served Qwen3.5 model that classifies through tool calls, while the rest of the dynamics stays fixed. Statistical model checking with MultiVeStA shows the expected ratio of happy agents rising from about 0.65 to about 1.0 within 60 steps in both variants, with point-wise t-tests finding no difference, yet the LLM version takes 8,917 seconds against about 2 seconds. The 0.8b and 2b models show semantic errors, wrong tool-call counts or stuck generation in unit-style checks, while 4b and 9b pass all of them.
Jag replaces coordinate-by-coordinate generation in visual grounding with direct regression of a bounding box in a single forward pass. Autoregressive multimodal models write a box as a coordinate sequence, so sequential decoding adds computation on top of the image-text processing that localization already needs. Jag attaches a lightweight regression head to Qwen3.5-0.8B, reads the last valid input-token state, and predicts all four continuous coordinates together under L1 and GIoU losses that jointly adapt the language backbone, visual merger, and head. On RefCOCO, RefCOCO+, and RefCOCOg it improves accuracy over its base model on every test split, with 65.97 ms mean single-request latency.