dots3-note Preview: A First Step Toward Long‑Term Agents for Real‑World Service

Model Specification

dots3-note Preview is the smallest model in the dots3 series with 280 B total parameters, 16 B activation parameters, and a 512 K token context window. It processes text, vision, and audio inputs and is released under the Apache 2.0 license on Hugging Face and GitHub.

TEMPO: Test‑time‑scaled Value Estimation with Macro‑step Policy Optimization

Long‑horizon, value‑free reinforcement learning suffers from extremely slow exploration (hours per episode) and sparse rewards that hinder credit assignment. Actor‑critic methods such as PPO improve efficiency but the critic’s fixed‑cost forward pass cannot exploit reasoning, reflection, or tool use to produce accurate value estimates for complex states.

TEMPO addresses these issues by partitioning a long episode into multiple macro‑steps, each consisting of several interaction rounds. At the end of a macro‑step the same agent switches from actor to critic and performs test‑time scaling: it runs an extended inference pass that reasons about the current state and estimates the expected remaining return. This estimate is used to update the policy before the episode terminates. During training the model learns both the action policy and the self‑evaluation capability.

Benchmark Results

On the ARC‑AGI‑3 benchmark, TEMPO raises the average score by 31.5 % relative to the baseline checkpoint and by 20.6 % relative to GRPO. The TEMPO model also reaches the same level with fewer steps. In the IMO 2026 competition a branch of dots3-note Preview generated proofs, recursively critiqued them with tool calls, and achieved a perfect 42/42 score.

Critic Analysis Example

In a “place‑knight” game the agent must place a specified number of pieces on a board while respecting hidden constraints. Two training trajectories run 64 steps without completing a new level. Trajectory B misinterprets the “piece‑attack” relation as the goal and continues searching under that false premise. Trajectory A correctly identifies the true conflict rule and discovers a near‑feasible layout. The critic reads both trajectories, incorporates privileged information about the hidden rule, and assigns markedly different value estimates to the two branches.

Recursive Self‑Evaluation

During TEMPO training the agent’s self‑evaluation ability exceeds expectations: even when the actor cannot solve a problem, the critic can distinguish a genuinely promising state from a misleading one and assign a significantly higher value. This empirical observation supports the hypothesis that, for the evaluated tasks, “evaluation is easier than generation.”

Real‑World Simulation Benchmarks

VibeSearchBench : 200 multi‑turn search tasks across 20 domains. Each task starts from a vague request and gradually discloses user intent. The agent may invoke search, visit, and code tools. Evaluation uses knowledge‑graph matching (Triplet F1) against a ground‑truth graph.

VibeLifeBench : 20 long‑horizon tasks across 10 domains, each spanning 20–30 phases with a simulated timeline. The environment models dynamic external factors (weather, prices, inventory) and provides 1 247 atomic checks for state consistency, tool execution, and final delivery. Scenarios include travel planning, medical appointments, and shopping, driven by persona‑based user simulators and asynchronous updates.

Memory‑Based Reasoning in ARC‑AGI‑3

dots3-note Preview demonstrates online learning and memory updates in the ARC‑AGI‑3 environment, where tasks can require thousands of interaction steps (40–50 h runtime). Example levels in a mirror‑coupled maze illustrate the reasoning pipeline:

Parse the ASCII frame to detect band‑aligned cells and token positions.

Construct a joint state graph for the two blue tokens (LEFT, RIGHT) with coupling rules (parallel vertical moves; mirrored horizontal moves).

Perform BFS on the joint state to find a sequence of moves that merges the tokens into the same cell.

Execute the plan in ≤10‑move chunks, verifying each step against the simulator and correcting divergence by re‑BFS from the actual state.

Level 2 solution moves the three markers to leaf nodes, opens all gated regions, and merges the tokens in 31 moves. Level 3 reveals a bug where the start assignment swapped LEFT/RIGHT columns; fixing the assignment restores a valid 31‑move merge path. Level 4 merges in 38 moves using pressure‑plate switches that open color‑specific gates. Level 5 (in progress) requires a 26‑move merge after unlocking a blocked corridor via a cursor‑based wall placement mechanic.

Evaluation Settings (selected)

Terminal‑Bench 2.1: 10‑hour timeout, temperature 0.7, top_p 0.95, max_tokens 81 920, 256 K context, 8 CPU, 16 GB RAM.

ARC‑AGI‑2: temperature 0.7, top_p 1.0, max_tokens 384 K.

ARC‑AGI‑3 (general harness): GPT‑5.5 evaluated with Codex, other models with Claude Code; all using a prior‑free harness.

ClawEval (General set): temperature 1.0, top_p 0.8, max_output 16 384 tokens, 262 K context, OpenClaw‑style system prompt.

WildClawBench: temperature 0.7, top_p 0.95, max_output 65 536 tokens, 320 K context, OpenClaw 2026.6.1 agent harness.

VibeLifeBench: OpenClaw harness with 256 K context, results reported as avg@3 on version v1.0.0.

Limitations

dots3-note Preview is a preview release; reinforcement‑learning training is not yet sufficient, leading to hallucination, multimodal balance, and stability issues. Deploying the model in real‑world applications requires additional infrastructure such as harnesses, connectors, data sources, security mechanisms, and product‑level refinements.