Latent Reasoning Directions (LRD)

Code and experiments are available in the project repository: bench-labs-org/Latent-Reasoning-Directions.

Update. The comparison has now been completed. Using the same 664 training examples, conventional fine-tuning reached 37.16% on ARC-Easy, while GPT-2 + LRD reached 36.15%. The gap is only 1.01 percentage points, despite LRD training just 854,208 parameters (0.69%) compared to 124.4 million parameters for full fine-tuning.

Important. These are still early results, not final conclusions. The experiments were performed on GPT-2 Small using only 664 reasoning examples. Larger models, larger datasets, and additional reasoning benchmarks are still needed to determine how well Latent Reasoning Directions scale.

Why Validation Loss Is Not Compared

The validation losses reported during training are not directly comparable.

Although these numbers appear very different, they measure different optimization objectives.

• GPT-2 + LRD trains a lightweight direction module to predict a latent reasoning vector using a regression-style latent-space objective.
• Standard fine-tuning trains every GPT-2 parameter using the language modeling objective (cross-entropy over tokens).

Because the models optimize different objectives and produce different training targets, their validation losses cannot be compared directly. Instead, both models are evaluated on the same downstream benchmark, ARC-Easy, making classification accuracy the meaningful metric for comparison.

Why "Latent Reasoning Directions"?

Unlike Chain-of-Thought prompting, LRD does not require generating reasoning tokens. The reasoning signal is compressed into a learned latent direction, making it substantially cheaper at inference time while remaining completely hidden from the output.

• Latent — the reasoning state is represented as a continuous vector rather than explicit text.
• Reasoning — the model learns an internal reasoning representation instead of being supervised with chain-of-thought traces.
• Directions — the vector acts as a direction in latent space, nudging the model toward a particular reasoning mode rather than encoding the reasoning process itself.

In other words, the model is not being taught how to reason step by step. Instead, it learns which reasoning mode to activate before solving the problem.

Does It Actually Improve Reasoning?

A clean latent space is encouraging, but visualization alone is not enough. The real question is whether these learned reasoning directions improve performance on tasks the model has never seen before.

To test this, I evaluated the model on the ARC-Easy benchmark. Importantly, the LRD model was not trained on ARC; it only saw a small collection of reasoning examples spanning mathematics, logic, language, commonsense, knowledge, and pattern recognition.

The improvement may appear modest at first glance, but on the full ARC-Easy evaluation set it corresponds to 57 additional questions answered correctly (859 versus 802).

Measured relative to the base GPT-2 model, this is a 7.1% improvement:

$$ \frac{36.15 - 33.75}{33.75} \times 100 \approx 7.1\% $$

Why This Result Matters

What makes this result interesting is the setting in which it was obtained.

• The direction module was trained on only 664 examples.
• The evaluation benchmark contains 2,376 science questions that the model never saw during training.
• The backbone is the original 124M parameter GPT-2 Small.
• No chain-of-thought supervision or reasoning traces were used.

Instead of teaching the model to produce reasoning steps, the goal was simply to learn a compact latent direction corresponding to different reasoning styles. During inference, that direction nudges the model toward a more appropriate reasoning mode before generating an answer.

Putting the Numbers in Context

The result lands roughly where I would expect. Since the model was never trained on science questions, it cannot compete with methods that fine-tune directly on ARC or distill large reasoning models. However, it consistently outperforms the pretrained baseline despite learning only from related reasoning tasks.

Comparing Against Standard Fine-Tuning

The obvious question after seeing the initial ARC-Easy improvement was whether the same 664 examples would be better used for conventional fine-tuning. To answer that, I trained every parameter of GPT-2 Small on exactly the same training set and evaluated it on the identical ARC-Easy benchmark.

Standard fine-tuning does achieve the best score, but only by 1.01 percentage points. The surprising part is the amount of training required to obtain that small gain.

The LRD model updates just 854 thousand parameters, while conventional fine-tuning updates all 124.4 million parameters of GPT-2. That is roughly a 146× reduction in trainable parameters, yet the final accuracy differs by only 1.01 percentage points.

LRD reaches 36.15% on ARC-Easy by training only 0.69% of GPT-2's parameters. Standard fine-tuning improves the score to 37.16%, but requires updating 100% of the model.

This shifts the interpretation of the earlier experiment. The original question was whether latent reasoning directions could improve a pretrained language model. They can. The more interesting question is now one of efficiency: how much reasoning ability can be recovered by training only a tiny fraction of the model instead of modifying every weight?

Takeaway

For me, this is the most encouraging result so far.

The latent reasoning directions are not merely producing visually separable clusters—they also translate into measurable improvements on an unseen benchmark. The gain is modest, but it demonstrates that a small reasoning module can transfer useful reasoning behavior across domains without requiring explicit chain-of-thought supervision.

There is still plenty of room for improvement. Larger backbone models, more diverse reasoning datasets, stronger direction encoders, and mixture-of-direction approaches are all promising next steps. Even so, this initial experiment supports the central hypothesis:

A compact latent reasoning direction can improve downstream reasoning performance without ever training on chain-of-thought traces.

One question I had while building the Latent Reasoning Direction (LRD) model was simple:

Did the model actually learn meaningful reasoning directions, or did it simply memorize labels?

To answer that, I projected every learned reasoning direction into two dimensions using t-SNE. Each point represents one learned direction from the training set. The left plot is colored by the six coarse reasoning categories, while the right plot uses the twenty-three fine-grained categories.

A Healthy Latent Space

The first thing that stands out is that the representation has not collapsed. Instead of forming one large cloud, the directions organize into several distinct regions. This agrees with the measured separation metrics, which reported a coarse-category kNN accuracy of roughly 94% and detected no representation collapse.

The coarse categories occupy their own areas of the embedding space. Mathematics forms clusters on the far right, Logic appears in the lower-left, Knowledge occupies the upper portion of the figure, while Pattern Recognition forms separate clusters below the mathematical reasoning directions.

Fine-Grained Structure

Looking at the right-hand plot reveals something even more interesting. Each coarse reasoning family naturally splits into smaller subclusters. Rather than learning only six prototypes, the model appears to organize different reasoning styles into a hierarchy.

For example, Knowledge separates into multiple neighboring groups, Logic forms its own curved manifold, and Mathematical reasoning breaks into several compact clusters. These groups remain close enough to indicate related reasoning patterns, while still being distinguishable from one another.

Smooth Transitions

Another observation is that many categories are connected by gentle curves instead of isolated circular blobs. Commonsense reasoning gradually transitions toward Language and Knowledge, reflecting their semantic relationship. These bridges suggest the model learned a continuous latent manifold rather than independent class centroids.

Takeaway

The experiments now suggest two separate conclusions. First, latent reasoning directions consistently improve over the pretrained GPT-2 baseline. Second, they recover most of the benefit of full fine-tuning while training only a tiny fraction of the model.

Training just 0.69% of GPT-2's parameters reaches within 1.01 percentage points of full-model fine-tuning on ARC-Easy.

Whether this scaling behavior continues on larger models and more challenging benchmarks remains an open question, but these early results indicate that latent reasoning directions are a promising parameter-efficient alternative to conventional fine-tuning.