AI Coding Live News

Looped language models reuse physical Transformer blocks across logical depth to compress parameter footprints, but existing approaches entangle recurrence placement, input injection, and recurrent-state topologies. CMU and Fudan researchers published Part I of 'Towards Looped Models Done Right' (arXiv:2610.23036). Introducing a Prelude-Recurrent-Coda sandwich architecture, an 8.0B-resident / 794M-active Huginn MoE trained on 500B tokens matches a 32.0B-resident / 3.2B-active feedforward reference across MMLU, GSM8K, and MATH500.

Zoom

Can LLMs reliably maintain spatial state representations across sequential operations? USC, UCLA, and CMU introduced Grid-Bench (arXiv:2610.22312, GitHub: grid-bench/Grid-Bench), an executable benchmark with 8,928 instances testing 2D/3D/4D grid state transitions across history lengths up to T=64. Evaluations across 6 frontier models show steep degradation: while GPT-5.6 Sol and Claude 5 Sonnet retain strong cell accuracy, open-weight and mid-tier models collapse to near-zero exact world reconstruction over long trajectories.

Zoom

Robotic world models struggle to predict physics when key environmental properties (friction, mass, contact dynamics) cannot be directly seen. Researchers from Stanford, CMU, and Toyota Research Institute introduced T2-z (arXiv:2610.22364, GitHub: T2-z-latent-optimization/T2-z). By freezing the world model and optimizing a compact environment latent z_e at test time from as little as a single interaction trajectory, T2-z halves physical error and raises real-world robot task success from 33.73% to 72.96%.

Zoom

Tackling reward hacking and benchmark saturation in coding agent environments, CMU researchers introduce Scalable Weak-to-Strong Verifier Hardening (arXiv:2610.15924)—combining an isolated verifier sandbox with an iterative hacker-fixer loop that slashes exploit success across Terminal-Bench, KernelBench, and SETA, and is now officially adopted by Terminal-Bench. Complementing this, Recursive Environment Scaling (RES, arXiv:2610.16397) recursively mutates execution dependencies across 294 tasks—exposing a 57.14-77.14 point drop on Claw-Eval-Live while boosting equal-budget SFT performance by 7.00-14.61 percentage points across all nine model-benchmark pairs.

Zoom

In production agentic loops, fast single-pass 'System 1' models are favored for latency-sensitive tool selection and interface actions. When these models fail, engineers cannot easily diagnose whether the agent misperceived environmental observations or misjudged its subsequent action. Researchers from Carnegie Mellon University (CMU) introduced a behavioral diagnostic benchmark across 17 fast models and frontier reasoning baselines, pairing next-action queries with independent factual probes about the state. The results expose a striking systemic pattern: the vast majority of erroneous actions follow completely accurate perception, with error concentration increasing in more accurate models. The true failure mode lies not in reading facts, but in synthesizing them into the next programmatic sub-goal.

Zoom

Humans solve complex spatial transformations by mentally simulating physical visual transitions, maintaining visual imaginations to ground multi-step reasoning. In contrast, existing Vision-Language Models (VLMs) tackle spatial questions almost entirely via text-token chains of thought, triggering acute hallucinations when predicting 3D rotations or spatial clearances. Researchers from Carnegie Mellon University and MBZUAI introduce WM-VLM (arXiv:2609.34826), integrating a lightweight world model branch directly into foundation VLMs to synthesize intermediate visual states during inference. The framework adopts a two-stage training paradigm: first training the model to foresee and render verifiable intermediate visual states, and then training it to ground downstream logical deductions upon those synthesized states. Evaluated on rigorous 2D and 3D mental rotation tasks, WM-VLM outperforms standard SFT baselines by up to 39.25 percentage points. Ablations prove that performance collapses when generated visual frames are removed or corrupted, establishing that internal world models provide indispensable spatial cognitive scaffolds.

Zoom

On-Policy Distillation (OPD) trains student LLMs on trajectories sampled from their own generation policies, providing robust alignment for mathematical and code reasoning. However, existing multi-teacher OPD frameworks restrict knowledge transfer purely to output token probability distributions, discarding rich semantic representations embedded in teachers' intermediate hidden states. Researchers from Carnegie Mellon University and Georgia Tech introduce Latent-MOPD (arXiv:2610.02381), the first representation-level multi-teacher OPD architecture for LLMs. Latent-MOPD integrates diverse specialists by coupling their output distributions with the internal representations used to generate them, without modifying teacher checkpoints. To coordinate disparate representations across specialists, Latent-MOPD selectively aligns late-layer targets, resolves channel mismatches through shared linear projections, and enforces domain-grouped gradient updates while gradually annealing hidden-state loss into token prediction. Across nine math, coding, and logic benchmarks, Latent-MOPD surpasses token-only and uniform-averaging baselines; matching the parameter budget of individual teachers, the distilled student outperforms the best domain teacher on a majority of benchmarks.

Zoom

Autonomous robot manipulation has long relied on specialized vision-language-action (VLA) models whose rigid training precludes zero-shot adaptation to novel environments, cutting them off from rapid foundation VLM advancements. Conversely, recent robotic agents rely heavily on fragile coding agents and auxiliary grounding models like SAM3, introducing severe latency and operational brittleness. Researchers from UIUC and CMU present MotorMind (arXiv:2609.38078), a direct robotic harness connecting general-purpose VLMs to deterministic robot control. By translating VLM mid-level actions into deterministic Cartesian primitives paired with asynchronous execution monitoring and background memory updates, MotorMind achieves 66.7% zero-shot success on LIBERO-PRO suites (versus 13.3% for prior baselines) and achieves 95% average success on a physical xArm6 robot under live human disturbances.

Zoom

Modern software engineering transcends isolated code edits: engineers run applications, interact with GUI interfaces, and visually inspect rendering feedback to diagnose errors and verify changes. CUA-SWE bridges coding agents and computer-use agents by introducing the first unified benchmark and environment for visual software engineering. Spanning four engineering domains, CUA-SWE challenges agents to interleave code/config modifications, terminal execution, GUI interactions, and screenshot inspections, with deterministic test suites evaluating behavioral correctness.

Zoom