Train the Decision, Not the Transcript: FSLR Targets the First Reasoning Choice

TL;DR for operators

A post-training team with a fixed budget can spend it on long worked solutions or on shorter labels aimed at the decision that determines the rest of the reasoning path. Wang and Zhang’s study¹ provides evidence that, for mathematical word problems, the second option deserves serious consideration.

The authors classify more than 90% of analyzed model failures as mistakes in selecting the relevant quantities, dependencies, or operation. Those errors remain dominant after conventional chain-of-thought supervised fine-tuning. Their alternative, First-Step Logical Reasoning (FSLR), trains the model only on that initial planning decision. Across the reported experiments, it generally beats matched full-trajectory fine-tuning, including under distribution shift, while using roughly 81-87% fewer supervised training tokens and approximately four- to sixfold less training time.

The practical lesson is not that reasoning traces are unnecessary. It is that supervision granularity is an allocation decision. If one early decision determines most downstream success, adding more tokens after that decision may be a poor use of the training budget.

The boundary matters: the paper tests mathematical reasoning on 4B-8B instruction-tuned models. It does not establish that first-step supervision will produce the same gains in software engineering, financial analysis, operational planning, or other production reasoning tasks.

More supervision can spend budget around the failure

Long chain-of-thought traces look like rich training data because they expose every intermediate step. But richness is only valuable when the additional tokens supervise the capability that is actually failing.

The paper’s diagnostic analysis places the failure earlier than arithmetic execution. Its category of logical relationship understanding covers mistakes such as choosing irrelevant quantities, linking conditions incorrectly, or selecting the wrong operation for the variables involved. More than 90% of the analyzed incorrect predictions fall into this class.

That attribution is not based on a single judge alone. The authors repeat the analysis with LLaMA-3.1-70B and Qwen2.5-72B as alternative judges, and a manual check of 100 random cases reports more than 95% agreement with the automated classifications. This does not make the taxonomy a mechanistic account of reasoning, but it strengthens the narrower diagnostic claim: within these experiments, structural mistakes dominate the observed failures.

That changes what training data should optimize. Full chain-of-thought supervised fine-tuning, or CoT-SFT, asks the model to reproduce the complete solution trajectory. The crucial variable-operation decision is present, but it is embedded among many later tokens describing calculations and subsequent steps.

FSLR trains the first planning decision, then lets inference proceed normally

FSLR removes most of that trajectory from the supervised target. For each problem, a teacher generates only the first planning step: identify the relevant variables and the operation connecting them, without completing the calculation or continuing the solution.

The restriction applies to training, not inference. An FSLR-trained model still produces ordinary complete solutions autoregressively at test time. This distinction matters because the method is not a one-step reasoning system. It is a proposal about where to concentrate the post-training signal.

The authors test three target models—LLaMA-3.1-8B-Instruct, Qwen2.5-7B-Instruct, and Qwen3-4B-Instruct—with supervision generated by large LLaMA and Qwen teachers as well as by the target model itself.

Focused supervision generally beats the full trace

The headline result is not merely that shorter supervision performs adequately. FSLR generally performs better.

The paper reports average gains over CoT-SFT of 3.2 percentage points in-distribution and 4.6 points out-of-distribution. With LLaMA-3.1-70B as the teacher, the average in-distribution improvement across the reported model and dataset settings is 4.8 points.

The out-of-distribution evidence is particularly relevant to the proposed mechanism. When models are trained on GSM8K and evaluated on ASDiv, SVAMP, MAWPS, TabMWP, and GSM-Hard, FSLR improves the average over matched CoT-SFT across all three target models and most teacher configurations. Reversing the setup—training on SVAMP instead—produces the same general direction.

The comparison with prompting also narrows the interpretation. Telling a base model to identify logical relationships at inference does not reproduce the fine-tuning result: the reported Zero-LP and Few-LP averages are 82.45% and 82.67%, while the corresponding FSLR configurations reach 87.08-88.90%. The result is therefore not explained by adding a planning instruction to the prompt.

The appendix tests whether the gain survives obvious alternative explanations

Several secondary experiments are better read as robustness checks than as separate claims.

The GSM-Symbolic result is the strongest of these for the paper’s central interpretation. If first-step supervision merely helped models reproduce familiar training forms, symbolic variation should erode the advantage. Instead, the gap widens substantially in the reported comparison.

The efficiency result turns supervision design into a budget decision

The training sequences make the cost difference concrete. On GSM8K, FSLR examples average roughly 27-52 tokens, compared with about 238-310 tokens for CoT-SFT. On SVAMP, FSLR uses roughly 34-46 tokens versus about 192-245.

The paper summarizes this as an 81-87% reduction in supervised training tokens, accompanied by approximately four- to sixfold lower training time.

For a team post-training compact models, that creates a different optimization problem. The relevant decision is no longer simply how many reasoning examples to collect. It is which part of each example deserves explicit supervision.

A defensible workflow would be:

1. classify failures by the decision that first sends reasoning off course;
2. identify whether one early choice predicts much of the downstream error;
3. construct a short supervised target for that choice;
4. compare it against full-trace supervision on held-out and perturbed tasks;
5. measure both accuracy and training cost.

The paper directly supports the token and training-time benefits in its mathematical setting. Cognaptus extends the implication one step further: if shorter labels are also cheaper to generate, inspect, or regenerate, the data pipeline may become cheaper as well. Human annotation cost, however, is not measured here.

The unresolved question is whether other domains have an equally compressible bottleneck

Mathematical word problems have unusually explicit structure. A small set of quantities and operations often determines the subsequent solution path. That makes the first planning decision a plausible compact supervision target.

Other domains may not behave this way. Coding failures can emerge from architecture, API assumptions, state changes, tests, or local implementation choices distributed across a long trajectory. Financial analysis may depend on evidence selection, temporal assumptions, accounting treatment, and uncertainty handling rather than one early operation. Operational planning can require repeated updates as constraints change.

The paper therefore supports a design hypothesis, not a universal recipe: before collecting long reasoning traces, test whether the task contains an early, high-leverage decision that can be labeled directly.

The strongest contribution of FSLR is not that less supervision is inherently better. It is that the amount of supervised text should not be confused with the concentration of the learning signal. In these mathematical reasoning experiments, training on the decision that structures the solution beats training on substantially more of the solution itself.

Cognaptus: Automate the Present, Incubate the Future.