Research Radar: A Coding Agent Drives a Two-Arm Robot to 39.7% on RoboDojo but Stalls at Tight Insertion, DLite Drops Draft History for Agent Workloads, and LEV Thinks Deeper by Looping Layers Instead of Tokens

This edition covers four arXiv preprints from October 14-16, 2026 that were listed on Hugging Face's October 19 Daily Papers.

HexaShell: a thin harness, a robot, and a coding agent (arXiv 2610.23479)

The authors ask "how far a coding agent can go when it controls a robot instead of a computer, and where it stops." HexaShell "only turns the agent's commands into robot motion": the agent drives a two-arm robot through one command line, sees three calibrated RGB cameras, and after each failure reads the log and writes or fixes small programs called tools -- keeping what it learns as code. The tools may use only what the robot observes, locating objects by triangulating camera rays or intersecting a ray with a known plane, and moving a wrist camera for a second view when needed. Tools are built on ten layouts of one seed per task, then frozen and tested on ten new layouts. Using Codex with GPT-6-Astra at medium effort, HexaShell reaches 39.7% success and a 48.0 progress score across all 42 RoboDojo tasks, which the authors report as a new state of the art. The finding that matters: "The limit is contact, not reasoning". Pushing reaches 92.5% and memory tasks 65.0%, but tight insertion only 8.3%, and when the agent fails "it usually sees the failure but cannot make the motion work." A companion SenseNova report (2610.22471) reaches a similar verdict from the other side: GPT-6-Astra "has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence," held back by expensive demonstrations and real-world time limits.

DLite: speculative decoding without the history (arXiv 2610.23787)

Agent workloads have long, constantly growing contexts and many turns, and the authors observe that existing draft models "suffer from up to an 80% degradation in acceptance length" under them, while keeping the draft's own KV cache costs system overhead. DLite reuses the target model's contextual hidden states and predicts locally, "without explicitly maintaining the historical context", with a light sequential head to restore dependencies inside each block. The authors report acceptance lengths of 2-6 and up to 70% better results than DSpark, the speculative decoding method SGLang 0.5.22 ships.

LEV: a decision model that loops (arXiv 2610.20951)

Decision models answer typed questions with one probability per option in a single forward pass, at a depth fixed in training. LEV "thinks deeper by rerunning its layers in a loop instead of generating tokens", reading the options after every loop so one checkpoint can answer at any depth. A 2.6B LEV rises from 63.2% at depth one to 76.2% at depth four on JevBench, and a small stopper that sets depth per question saves 15-24% of compute for under one accuracy point.

TreeOPSD: credit assignment without a critic (arXiv 2610.23181)

For long-horizon agent RL, TreeOPSD branches alternative continuations from shared turn prefixes, backs up terminal rewards into turn-level credit, and distils toward a bounded teacher distribution, with no value network or process reward model. It reports a 4.97-point gain over GRPO for Qwen3-30B-A3B-Thinking-2507 in a deep-research setting.

Analysis: HexaShell's split result -- strong at perception and planning, weak at contact -- is a useful map for anyone pairing a language-model agent with hardware: let the agent write and debug perception code, and hand contact-rich motion to a learned low-level controller, which is the System 2 / System 1 split the SenseNova report also recommends. DLite and LEV point the same way for inference: agents need cheaper extra steps, whether that means drafts that do not drag the whole history along or answers that buy depth with loops instead of tokens.