How do LLMs do CoT reasoning internally?
In our new #ACL2026 paper, we show that reasoning unfolds as a structured trajectory in representation space. Correct and incorrect paths diverge, and we use this to predict correctness before the answer and correct errors mid-flight.
1/
In our new #ACL2026 paper, we show that reasoning unfolds as a structured trajectory in representation space. Correct and incorrect paths diverge, and we use this to predict correctness before the answer and correct errors mid-flight.
1/
Each reasoning step occupies a distinct region in representation space, and these regions become increasingly linearly separable with layer depth. This structure generalizes across tasks and answer formats.
2/
2/
Trajectories for correct and incorrect reasoning start out nearly identical but diverge systematically at late steps. Late-step trajectory features predict final-answer correctness with AUC 0.87 - before the answer is generated.
3/
3/
This step-specific organization exists across training regimes and is already present in base models. From this perspective, reasoning training doesn't introduce new representational organization - it accelerates convergence toward a termination-related subspace.
4/
4/
Unconditionally injecting tokens like "Wait" often hurts accuracy (up to -36%, especially for fewer-step problems). Instead, intervene only when failure is predicted! Using our mid-reasoning correctness signal on just ~12% of examples yields stable gains.
5/
5/
We further introduce trajectory-based steering: when an ongoing trajectory drifts from the ideal path, we apply low-rank corrections to nudge it back. Most effective on harder math problems - +7.6pp on 6- and 7-step questions, with 97%+ preservation of correct solutions.
6/
6/
Reasoning length is also controllable. Steering hidden states toward the termination subspace shortens reasoning; steering away extends it. At moderate strengths this works as a smooth knob with minimal accuracy changes - push too hard and the model enters repetitive loops.
7/
7/
📢 Accepted to #ACL2026 Main Conference!
Thanks to all collaborators at Microsoft: Hang Dong, Bo Qiao, Qingwei Lin, Dongmei Zhang, and Saravan Rajmohan.
Paper: arxiv.org/pdf/2604.05655
Website: slhleosun.github.io/reasoning_traj/
Code: github.com/slhleosun/reas…
8/
Thanks to all collaborators at Microsoft: Hang Dong, Bo Qiao, Qingwei Lin, Dongmei Zhang, and Saravan Rajmohan.
Paper: arxiv.org/pdf/2604.05655
Website: slhleosun.github.io/reasoning_traj/
Code: github.com/slhleosun/reas…
8/
• • •
Missing some Tweet in this thread? You can try to
force a refresh
