AI FRONTIER
Signals worth understanding before they become obvious.
Why Sub-4-Bit Vision CNNs Collapse and Where Mixed-Precision Recovery Stops
The paper systematically analyzes quantization collapse across standard Vision CNNs, including ResNet and EfficientNet backbones, evaluated on the full 50,000-image ImageNet-1k validation set, and introduces a multi-step recovery waterfall.
- Collapse is highly non-linear and concentrated in late-stage pointwise projections and depthwise layers.
- Keeping less than 5% of critical layer activations in 8-bit precision recovers up to 92% of the accuracy lost during aggressive sub-4-bit compression.
- On standard ARM and x86 CPU architectures, standard runtime software bit-unpacking introduces a 1.48x to 21.5x execution slowdown depending on batch size.
For edge vision builders, sub-4-bit weights cut storage rather than throughput, so mixed-precision recovery should target late-stage pointwise projections and depthwise layers while measuring unpacking overhead.
thinkingmachines/Inkling-2-Mini: a 26.4B multimodal looped transformer with open weights
Thinking Machines Lab released Inkling 2 Mini, a natively multimodal looped transformer with 26.4B total and 3.4B active parameters, under Apache 2.0 with open weights. It accepts text and image inputs and supports a context window of up to 1M tokens.
- 26.4B total parameters and 3.4B active; the middle 12 layers run three times with shared weights, giving 48 layer applications per token and 5.4B effective active parameters.
- Architecture is a 24-layer decoder-only sparse MoE that routes each token to 8 of 128 experts per layer, with hybrid local and global attention.
- Distributed as a single checkpoint of about 20GB in mixed NVFP4 and MXFP8 precision, with supported local inference on 1x NVIDIA DGX Spark and Apple silicon devices with 32GB or more memory.
- The model card reports SWEBench Verified (bash only) 79.2%, SWEBench Pro v2 (Hard) 77.9%, MMMU Pro Standard 10 74.9%, and OmniDocBench v1.6 92.8%.
Builders get a multimodal open-weights model that fits 32GB-class local hardware, but the model card states it is less capable than Inkling 2 Small on factual recall and may take destructive agentic actions without confirmation, so sandboxi...
Hugging Face introduces harness adapters: adapting a model to your agent harness with RL
Hugging Face published a reusable recipe that trains one LoRA adapter per agent harness for Qwen3.8-27B with RL inside Pi, OpenCode and Hermes Agent, then merges the three LoRAs into a single super adapter that runs unchanged under all three harnesses.
- Each adapter was trained on NVIDIA 2×B200 GPUs, with one GPU for the vLLM server and one for the training process.
- The final data mixture combines 750 tasks from harness-adapters RL environments with 300 tasks from termgrade.
- The harness-hub/agentic-coding-envs dataset holds 871 tasks from 35 repositories, of which 171 form the qualified subset for GRPO training.
For agent builders, this shows that RL fine-tuning inside the harness you actually deploy can raise pass rates while cutting token usage, without swapping the base model.
Training Language Models for Introspective Oversight
The paper introduces introspective oversight: a separate third-person observer role is added to the model to answer questions about the assistant, rewarded for answers verified against the assistant's actual behavior. On five held-out oversight evaluations, an observer trained on Qwen3.6-27B obtains a mean AUROC of 0.9...
- An observer trained on Qwen3.6-27B obtains a mean AUROC of 0.907 on five held-out oversight evaluations.
- Baselines: honesty training 0.864, transcript monitoring 0.828, activation oracles 0.796.
- The paper reports that training an observer makes the assistant itself easier to monitor, with every baseline method improving on average on the fine-tuned assistant.
- The paper is 31 pages with 13 figures and 8 tables, submitted on 17 Nov 2026.
For builders, this suggests a trainable observer role, trained separately from the assistant solely for accuracy, as an oversight path beyond external transcript monitoring and activation decoders.
Watch Where You Linearize: Predicting and Correcting Model-Merging Interference
The paper shows that a first-order directional derivative of task A's outputs along task B's vector can predict interference, but only when evaluated at the receiver model, and builds LR-Merge on this to optimize merge coefficients via label-free convex teacher matching.
- At the receiver model, the derivative forecasts nearby merges per task and per example, halves the accuracy error of a no-change baseline, and identifies most examples whose predictions flip.
- Full multi-task merges require two evaluations along the merge path; a single evaluation is insufficient.
- At the shared pretrained backbone, the derivative is nearly orthogonal to realized changes in the vision experiments and only partly predictive in LoRA language-model pilots.
Builders can evaluate the directional derivative at the receiver model to predict merge interference cheaply and use LR-Merge's label-free convex teacher matching instead of expensive coefficient search.
Where the Tokens Go: Heat, Entropy and Error in Agentic Coding
The paper introduces a theory of software thermodynamics that counts tokens generated by a coding agent as cost and divides that cost into work and heat. It tests the theory by having an agent repeatedly make and undo repository changes while counting tokens of every model call.
- The paper reports that about 80 per cent of the tokens of a successful run are generated in model calls that leave the requested change where it was, which the authors call heat.
- The paper reports that the free energy of a change, estimated by analogy with the equality of Jarzynski, is 14 to 22 per cent of the mean cost of a successful coding run.
- The paper reports that coding costs 1.1 to 1.5 times as much as uncoding.
- The paper reports that the spread of cost over repeated cycles defines a software temperature, which rose with the sampling temperature of the model.
The framework gives builders a way to split agent token spend into useful work and wasted heat so they can target model calls that do not advance the requested change.
RFBench: A Simulator-Grounded Multimodal Benchmark for Passive RF Layout Design and Understanding
The paper introduces RFBench, a simulator-grounded dataset and benchmark for evaluating multimodal language models on passive RF layout design and understanding across seven tasks.
- The dataset contains 10,000 two-port planar layouts simulated using ADS Momentum.
- Each layout includes editable geometric primitives, a raster image, a 16x16 occupancy grid, complex S-parameters at 201 frequency points, a structural category, and an RF response label.
- The paper reports that RF-specific fine-tuning increases T1a structural validity from 0.0% to 83.2% and the T3 primitive-input within-2-dB rate from 17.7% to 62.8%.
RF-specific fine-tuning substantially improves structural validity and response prediction, but the reported results show considerable room remains for generating electromagnetically faithful layouts.
Global Importance Estimation for KV Cache Eviction
The paper introduces PRKV, a training-free KV eviction method that estimates multi-hop global token importance by applying Personalized PageRank over attention-induced transition graphs, enabling importance propagation beyond recency or instantaneous attention without modifying the model.
- PRKV is training-free and does not modify the model.
- The method reduces peak memory usage with low end-to-end overhead.
- The paper was accepted at NeurIPS 2026, by Seojin Kim and Noseong Park.
For builders of long-context inference, PRKV indicates that replacing recency or instantaneous attention scoring with multi-hop global influence estimation can preserve task performance at high compression ratios while lowering peak memory.