An Experiment: Making the Student Skip the Work
Imagine a math class. A student solves a problem, writing out a long, elegant derivation before reaching the correct answer. The teacher is satisfied—until he runs an experiment: he asks the student to skip the derivation and write the answer directly.
The result: the student's accuracy when writing answers directly is nearly identical to accuracy with a full derivation.
What does this mean? The derivation may not be the "path to the answer" but rather a "justification written after the answer was already known."
This is exactly what Jingyu Liu et al. (Renmin University + ByteDance) discovered in the paper *Where Do CoT Training Gains Land in LLM based Agents?* (arXiv:2606.26935). Through a series of carefully designed experiments, they show: the gains from CoT training mainly do not live in the CoT.
The Core Question: What Does Training Actually Improve?
Chain-of-thought (CoT) reasoning is standard equipment for LLM agents today. The model generates a reasoning chain first, then outputs an action. The mainstream narrative: CoT training teaches models to "reason," and longer, more complex chains mean smarter models.
But prior research on "unfaithful CoT" has shown that models may already know the answer from the prompt before generating CoT—the CoT is just post-hoc rationalization. If that's true, what does CoT training actually improve? Is it making models better at "revising actions through reasoning," or better at "directly predicting actions from the prompt"?
The paper designs a remarkably clean controlled experiment: comparing two decoding modes.
- CoT action: the standard mode—the model reasons first, then outputs an action.
- Prompt action: an
<action>tag is pre-filled at the start of the response, forcing the model to skip reasoning and output an action directly. - Roughly 80% of attention mass falls on prompt tokens, not CoT tokens.
- Even examining top-K highest-attention positions, prompt tokens dominate—this is not merely the mechanical effect of "longer prompt means more attention."
- Gradient decomposition along the value path shows prompt tokens also receive a larger share of gradient signal.
If CoT training genuinely improves "reasoning ability," then the advantage of CoT action over prompt action should grow with training—models should increasingly rely on reasoning to correct their actions. Conversely, if training is (or mainly is) improving direct prediction, accuracy in both modes will rise in parallel, with CoT's advantage unchanged.
Three Pieces of Evidence: The Gains Land on the Prompt Side
Evidence 1: Parallel Improvement
On ALFWorld (household tasks), ScienceWorld (scientific reasoning), and BFCL (function calling), the authors train Qwen3-8B and Llama models with SFT and RL, comparing the two decoding modes across checkpoints.
The result is clear: prompt action and CoT action accuracy rise in parallel. Models improve with reasoning, and improve equally without it. More tellingly, the frequency of both modes arriving at the same action (prompt/CoT consistency) also rises—models increasingly give the same answer whether or not they reason.
This is not what "reasoning getting stronger" should look like. If reasoning were key, the marginal benefit of reasoning should grow. It doesn't.
Evidence 2: The Gap Stays Flat in Online Evaluation
One might object that parallel offline accuracy gains are just a memorization effect. So the authors ran online interactive evaluations on unseen tasks.
Figure 3(a) shows the trajectory success rate gap between CoT action and prompt action stays flat throughout training. Figure 3(b) uses GPT-5.4 as a judge to compare action quality at the same decision points—the win rate of CoT over prompt also remains stable.
Training did not make CoT more "useful." It made both modes better together.
Evidence 3: Even Conflicting Reasoning Chains Can't Move Later Models
This is the most elegant strike.
The authors ran a "conflicting reasoning chain" experiment: they replaced the model's original reasoning chain with one supporting a different action, then observed whether the model's final action followed the original prompt or the injected chain.
Result: later checkpoints increasingly preserve the prompt-implied action, even when the reasoning chain says otherwise.
This directly refutes the hypothesis that training strengthens the use of reasoning. If training reinforced the reasoning-action causal link, replacing the chain should more easily change actions. The opposite happens—training makes models more "stubborn" about following the prompt, and the reasoning chain's influence declines.
In the paper's words: later checkpoints are less likely to revise the action in response to CoT.
Why Does the Prompt Have a Structural Advantage?
The authors go beyond behavioral evidence and offer a mechanistic explanation.
In agent settings, prompts contain task instructions, interaction history, and environment feedback—usually far longer than the CoT. This length asymmetry creates a structural attention hegemony:
This is a structural "lock-in": the longer the prompt, the more marginalized the CoT. And agent prompts are naturally long—they must hold the entire interaction context.
A Counterintuitive Intervention: Supervise Actions a Little Less
If the problem is that "action supervision mainly reinforces the prompt→action shortcut," the solution follows naturally: mask action-token supervision on a portion of training samples, optimizing only the CoT span.
The method: randomly select k% of samples, compute loss only on the CoT portion, and mask the action loss on those samples; other samples keep standard supervision. The authors call this "reduced action supervision," with k=0.3 performing best.
The intervention is simple enough to be suspicious, but it achieves two things at once:
1. Improved OOD performance: on ALFWorld and ScienceWorld, most model-environment combinations improved (Figure 6). Gains were smaller on BFCL—consistent with the prediction that "the stronger the prompt shortcut, the greater the benefit of the intervention," since BFCL's function-calling prompts are relatively short. 2. The CoT-prompt gap reopens: after the intervention, CoT action's advantage over prompt action grows (Figure 7), showing CoT regains some "say" in changing actions.
On Qwen3-8B, SFT + reduced action supervision combined with DPO raised average performance across three environments from 0.61 (FRODO baseline) to 0.67 (Table 1).
The intervention is counterintuitive because it deliberately gives up some action supervision—under the intuition that "actions are the final goal," this looks like weakening the objective. But the effect is the opposite: precisely because standard training sends action supervision down the wrong path (reinforcing the shortcut), giving slightly less action supervision helps models learn better.
What Does This Mean?
"CoT Training" May Be a Misnomer
If the main gains from training land on prompt action, then the name "CoT training" itself is misleading. A more accurate description: we are training models to better predict actions directly from prompts, with CoT as a byproduct.
This closes the loop with unfaithful-CoT research: CoT is unfaithful at inference time (the model already knows the answer) and unfaithful in training (training doesn't improve reasoning itself)—CoT is never the protagonist.
Reasoning Is the Witness, Not the Detective
An analogy: the prompt is all the physical evidence at the crime scene; the CoT is the witness in court. We thought training the witness would make the detective better at solving cases, but the detective increasingly relies on physical evidence—the witness merely repeats what the evidence already implies. The more training, the more the witness sounds like a tape recorder, and the less the detective listens (the conflicting-chain experiment).
A Warning for AI Safety
If CoT mainly "shows the work" rather than "does the work," then CoT-based interpretability and CoT-based alignment both rest on a fragile assumption—that CoT faithfully reflects the model's decision process. This paper provides stronger evidence: CoT is unfaithful not only at inference time, but becomes *more* unfaithful with training. Training doesn't make CoT more "genuine"—it makes it more "decorative."
Engineering Takeaway: Less Is More
Reduced action supervision is a directly usable technique. If you're training agents—especially long-context agents—supervising only CoT on 30% of samples and skipping action supervision may outperform full action supervision. The implementation is trivial: just mask the action tokens' loss on a subset of samples.
Limitations and Open Questions
The paper candidly notes: prompt action is only a behavioral proxy for "how much action information is recoverable from the prompt," not a direct measurement of internal computation. Some judgments rely on GPT-5.4 as a judge, which may introduce evaluation bias. Moreover, the paper doesn't fully dismiss CoT's role—it claims that "the gains of standard CoT supervision should not be read as evidence of improved CoT revision ability," not that "CoT is entirely useless."
A deeper question remains: if training doesn't improve reasoning, what will? The paper's intervention (reduced action supervision) only makes CoT "somewhat useful again"; it doesn't make CoT the dominant path. Training methods that genuinely improve reasoning may need structural redesign—for example, balancing gradient flow between CoT and action, or architecturally preventing the prompt→action shortcut.
Conclusion
This paper does what science should do but rarely does: negate a widely assumed default.
The mainstream narrative says "CoT training improves reasoning." This paper says: in the environments measured, the evidence does not support that claim. Training improves direct prompt→action prediction; CoT's relative advantage stays flat or declines; conflicting reasoning chains increasingly fail to move actions—none of this is what "reasoning getting stronger" should look like.
Feynman said: "If you can't explain a concept in simple terms, you don't truly understand it." This paper's variant might be: if the model can give equally good answers without reasoning, it probably wasn't reasoning in the first place.
Sometimes scientific progress isn't discovering a new continent—it's discovering that the continent drawn on the map doesn't actually exist.
---
Paper: *Where Do CoT Training Gains Land in LLM based Agents?* Authors: Jingyu Liu, Zhiwen Wang, Yuxin Jing, Huanyu Zhou, Yong Liu Institutions: Gaoling School of AI, Renmin University + ByteDance arXiv: 2606.26935 Code: Not released