October 30, 2026News · Research

Introducing Causal Reflection

We are releasing initial benchmark results for a new model system architecture that enables companies to own their intelligence without fine-tuning, post-training, reinforcement learning, or routing.

Today, enterprises face a cost-reliability tradeoff: either pay for closed (frontier) models like Claude or GPT on hard tasks, or accept weaker performance and reliability from open models. Fine-tuning, RL, and routing add cost and complexity without solving the deeper problem: agents stay static while your workflows change.

Based Models began with the hypothesis that many of the failures that force enterprises onto frontier models are failures of execution, not knowledge. More directly, agents lose track of state, make incorrect assumptions (hallucinations, lack of context…), fail to recover from tool errors, and continue operating after plans break. For example, Cheng et al. (2026) found that many tool-use failures do not stem from a lack of knowledge, but from a failure to convert that knowledge into action, revealing a "knowing-doing gap" in agent execution.

Our research demonstrates that many of these failures can be corrected without changing the underlying model weights. This capability is unlocked by our core technology, Causal Reflection - a reliability layer that detects when execution diverges from expectations and helps agents recover before errors compound. The result is frontier-level performance and higher reliability on open models, so you can keep your stack open-source and own your intelligence.

With open models, Causal Reflection does three things:

• Boosts cheaper, mid-tier models to frontier-level performance.
• Increases reliability on multi-step task execution.
• Continuously adapts and improves as your workflows and environment evolve.

Results

We measured performance across three benchmarks using GLM 5.3 Flash as the base model.

Terminal-Bench 2.1 (boosted performance on coding and terminal tasks) [1]

Terminal-Bench 2.1 measures an agent's ability to autonomously operate inside a real, sandboxed command-line environment to complete complex engineering, system administration, and technical tasks. In our environment, GLM 5.3 Flash achieves a 79% score, representing a capable and cost-efficient setup that reliably handles standard workflows but falters on complex edge cases. With Causal Reflection, GLM 5.3 Flash is boosted to the frontier with near-perfect accuracy across highly complex, multi-step system operations and advanced debugging challenges.

τ²-bench Airline (multi-turn customer service reliability, Pass^1 to Pass^4) [2]

τ²-bench Airline measures an agent’s reliability in multi-turn customer service conversations, where it must follow domain policy and use tools to resolve requests like flight changes and cancellations. Pass^k scores require the agent to succeed on the same task in all k attempts, making Pass^4 a strict test of consistency. In our environment, GLM 5.3 Flash scores 82% at Pass^1 but falls to 66% at Pass^4, showing a capable model whose reliability erodes across repeated runs. With Causal Reflection, GLM 5.3 Flash rises to 88% at Pass^1 and holds 76% at Pass^4, outperforming Opus 5 at every level while maintaining a stable performance spread (i.e. matching frontier reliability) and widening its lead over the base model.

Continual Learning Bench (adaptation across changing task sequences) [3]

Continual Learning Bench measures whether an agent improves from experience, evaluating systems over sequences of related, expert-validated tasks in domains like software engineering, data science, and strategic modeling, where the system is expected to change as it works. Normalized Reward captures overall task performance, while Normalized Gain compares each system to its stateless version to separate what it actually learned from prior tasks from its raw capability. On the leaderboard, the next-best entry, in-context learning with Claude Sonnet 4.6, reaches a Normalized Reward of 0.1960 and a Normalized Gain of 0.2410 at an average cost of $30.43. With Causal Reflection, GLM 5.3 Flash reaches a Normalized Reward of 0.4245 and a Normalized Gain of 0.4856 at an average cost of $2.32, more than doubling the #2 entry on both metrics at over 10× lower cost, and showing that its improvement comes from learning rather than raw capability.

Join Our Closed Beta

The first deployment of Based Models will be via API. The API adapts online to your workload and improves from the feedback it receives. We manage session artifacts on the backend; you get boosted performance, reliability, and observability (in your dashboard) out of the box. There is no separate harness or external memory store to wire up.

We are opening a limited beta. Sign up here for the waitlist.

Notes:

[1] Terminal bench 2.1 official board. Custom board hosted on Harbor Hub, click to see our runs and trajectories.

[2] τ²-bench Airline official board. Our submission as leaderboard PR. Custom board hosted on Harbor Hub.

[3] Continual learning bench official board. Our submission as leaderboard PR.