Today's theme is System One decision models: SO-Arena shows the most accurate model is not the most reliable, and decision models can be good teachers for agents. Google launched Gemini 4 Hydrogen and Alibaba opened a Qwen4-Max preview. Agent studies expose wasted steps after a fix and the weight of harness design, while voice brings Duo (audio-visual full-duplex) and DialInstruct-TTS. The window was extended to about 3 days because arXiv listings had not updated at run time.
Large Language Models
1
SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models
SO-Arena benchmarks System One models, which answer in a single pass by choosing among options instead of generating long reasoning. It spans 11 domains and five capability families and evaluates 21 models. Beyond accuracy, it measures reliability through decision stability under controlled option-set changes and recognition of missing correct answers. Gains from scale and extra reasoning vary sharply by domain. Notably, the best-performing System One model ranks only seventh in reliability.
Why it mattersDecision models (such as Gemini Decision 1, launched this week) are becoming a product category, and SO-Arena is one of the first benchmarks to test them beyond accuracy. The finding that top accuracy does not imply reliable decisions matters for anyone using them as routers, judges or classifiers. It was the top paper on Hugging Face Daily Papers for Nov 13.
Factored One-sided Adam-Moment for Practical and Scalable SOAP
FOAM is a lightweight variant of the SOAP optimizer for transformer training. It preconditions only the input side, dropping the output-side Kronecker factor and one eigendecomposition, and replaces SOAP's dense rotated second moment with a factored row-by-column estimator. As a result it uses half of SOAP's optimizer state and runs at Muon-class wall-clock speed. From 124M to 760M parameters it matches strong SOAP and Muon baselines in validation loss and stays robust across a three-fold learning-rate range. Code is released.
Why it mattersSecond-order-style optimizers such as SOAP and Muon are the main challengers to AdamW for LLM pretraining, but SOAP's memory cost limits adoption. FOAM halves that state while keeping quality, though results so far stop at 760M parameters, so behaviour at billion-parameter scale is the open question.
Actor, Judge, or Teacher? Understanding Decision Models in Long-Horizon Agents
The paper studies decision models, which return a choice with probabilities, in three roles inside long-horizon agents: actor, judge of completed actions, and teacher that scores actions for RL. On ALFWorld, all six tested models (three decision models, three general LLMs) were pulled away from a required subgoal by an easier one when acting. As judges with the full trajectory, the same models credited steps toward that subgoal, and asking whether each action helped separated useful actions better than asking whether the task would succeed. Used as a teacher with GRPO, a Qwen2.5-1.5B policy rose from 72.8% to 82.9% on ALFWorld and from 56.8% to 71.4% on WebShop. Code is released.
Why it mattersIt gives a concrete recipe for using cheap decision models as step-level reward signals in agent RL rather than as the agent itself. The finding that hindsight judging beats foresight acting suggests a practical division of labour for training long-horizon agents.
Termination Lag: Coding Agents Solve Early but Stop Late
The authors define termination lag as the number of steps between a coding agent's first correct patch and the end of its run, across five models, two agent frameworks and three SWE-bench datasets. 38–82% of successful runs reach a correct solution on their first source-code edit. After that first correct patch, remaining steps make up a median 35–37% of successful runs for Qwen3-Coder and 54–60% for Laguna-XS.2 on SWE-bench Verified and Multilingual. Most of these steps go to re-verification and repository exploration. In 1–8% of cases the continued activity makes the agent undo its correct patch and finish with a wrong solution.
Why it mattersA third to over half of a successful coding agent's steps can come after the fix is already in place, a direct cost and latency target. Better stopping policies could cut agent spend substantially and even raise accuracy by avoiding self-inflicted regressions.
JustRUA: A Controlled Study of GPT-6 Robot-Use Agents
JustRUA is a controlled study of design choices for robot-use agents built on GPT-6 Astra, comparing the harness, action interface, experience reuse and scene geometry within one framework. A lean harness that keeps the full episode history and stops each motion early performs well. Experience reuse changes how the model acts but does not reliably help, while scene geometry helps most, even when depth is predicted from RGB alone. The JustRUA harness, trimmed from the official L3 harness, needs no learned policy, self-improvement loop or task-specific tuning. It raises the RoboDojo simulation score from 29.7 to 50.3; code is released.
Why it mattersFor frontier-model robot agents, harness and perception design can move scores as much as a model upgrade. The ablations give practitioners a clear priority: give the model scene geometry before investing in memory or self-improvement loops.
Hydrogen is Google's fast, cost-efficient model for agentic coding and everyday work, complementing the frontier Gemini 4 Carbon and Argon. It has four thinking levels (low, medium, default high and a new xhigh), a 1M-token context, and costs $1.50 per million input tokens and $7.50 per million output tokens. Google reports 51 on the Artificial Analysis Intelligence Index at high (54 at xhigh, versus 50 for Argon), 57% on Terminal-Bench (Gemini 3.8 Flash: 18.2%) and 58.8% on HLE-Diamond. It finishes tasks about 2.5x faster than 3.8 Flash and is available in the Gemini API, AI Studio, Antigravity, the Gemini app and Search. Figures are vendor-reported; a model card is published.
Why it mattersA mid-priced model matching or beating Google's own frontier Argon on aggregate indices compresses the price-performance curve for agentic coding. It completes the Gemini 4 lineup days after Carbon and raises pressure on competing fast tiers.
Alibaba opened a preview of Qwen4-Max, the first Qwen4 model developers can call via API. It accepts text, images and video, has a 1M-token context and up to 393,216 output tokens shared by thinking and answer, with thinking on by default via a reasoning_effort setting (low, medium, xhigh). It supports function calling, web search and structured outputs, and is compatible with the OpenAI Chat Completions and Responses APIs and the Anthropic Messages API. The preview runs only through Token Plan from Nov 11 to the end of Nov 25, 2026 (UTC+8) and will then be retired. Alibaba has not published benchmarks, parameter counts, license or weights.
Why it mattersIt is the first look at Alibaba's next-generation flagship and signals that the Qwen4 family is close. Without benchmarks or a model card its real standing is unknown; watch for the full release after the preview window.
IBM introduced LLMON (LLM Object Notation), a markup language that makes structure and meaning explicit to LLMs: delimiters are reserved special tokens and unique IDs bind each instruction to the data it acts on. Two experimental previews, granite-4.1-3b-llmon-preview and granite-4.1-8b-llmon-preview, are built on Granite 4.1, with converters between JSON and LLMON. On the OPI prompt-injection benchmark IBM reports LLMON avoids 47% more attacks than the best frontier model tested and 95% more than the best open model, with similar trends on ASB and StruQ. Standard benchmark quality is on par with Granite 4.1 Instruct. These are previews for evaluation, tested so far only against static attacks.
Why it mattersPrompt injection is the main security blocker for tool-using agents, and LLMON tackles it at the token and format level rather than with filters. If results hold under adaptive multi-turn attacks, structural separation of instructions and data could become a standard part of chat templates.
DialInstruct-TTS: Interactive Expressive Speech Generation via Multi-Turn Dialogue with Fine-Grained Temporal Control
Instruction-based TTS systems set one style for a whole utterance, so small changes require rewriting the instruction and regenerating everything. DialInstruct-TTS lets users refine speech through multi-turn natural-language dialogue. A Qwen-based planner turns the dialogue history and latest instruction into a complete TemporalPlan, and a temporal acoustic control module adjusts emotion, speaking rate, volume and phonation per segment. A reference-latent revision mechanism keeps non-target regions unchanged, so specific attributes or time spans can be edited. Subjective evaluations show better instruction following, local style realization and preservation of untouched regions; the abstract reports no quantitative scores.
Why it mattersIterative, localized editing is how human voice directors work, and this is a step toward conversational audio editing in production TTS. The evidence is subjective only so far, so independent quantitative evaluation is needed.
Duo: A Reciprocal World Model for Full-Duplex Interaction
Duo v1 is a full-duplex world model that continuously perceives audio-visual input while responding to users in real time. The authors propose the Reciprocal World Model framework with two capabilities: Intent, which reasons about the interaction and when to participate, and Realization, which generates coherent audio-visual behaviour. Its three main designs are autoregressive modelling for continuous interaction reasoning, synchronized tokens on a unified timeline with attention masks that block future inputs, and dual rollout distillation to reduce distribution shift. The system supports spontaneous conversation and interruptions, and lets users show real objects, reshape virtual scenes or play audio-visual games without mouse or keyboard. The abstract gives no benchmark numbers.
Why it mattersDuo extends full-duplex interaction from audio-only to joint audio-visual perception and generation, pointing toward always-on embodied conversational agents. It drew strong attention on Hugging Face Daily Papers, but without quantitative results its latency and robustness remain to be checked.
