BIGAI GradCuit lifts LLM test-time reasoning by 6.6 points

A team from Beijing's BIGAI, Peking University, the Chinese Academy of Sciences and BUPT posted a test-time reasoning method called GradCuit that squeezes more accuracy out of already-trained language models by editing their internal activations at inference. Their paper on Hugging Face reports an average 6.6 percentage-point gain over standard chain-of-thought across five open backbones and three reasoning benchmarks, without touching model weights.

The trick is where the optimization lives. Prior work such as LatentSeek inserted optimizable latent variables at the transformer's output side and backpropagated through decoded tokens to update them, which creates what the authors call a credit-assignment problem: the update signal for each latent is blurred through the decoder. GradCuit instead injects those learnable latents into an intermediate hidden layer, so the model's own self-attention acts as a routing circuit that carries gradient signals directly from every generated token back to every latent. Hence the name, gradient through circuit.

The numbers are consistent rather than a blowout. GradCuit hits 64.5% average accuracy across 30 backbone-benchmark-format settings, against 62.1% for the strongest competing method, and wins the best individual result in 23 of 30 configurations. Gains are largest on MATH-500, where it beats LatentSeek by 2.8 to 8.9 points depending on the answer format. More telling for practitioners: across seven learning-rate settings the accuracy standard deviation drops from LatentSeek's 1.53 to 0.82, and even a random-walk version of GradCuit, with reward-derived gradients replaced by Gaussian noise, hits 60.6% average, slightly ahead of reward-guided LatentSeek at 60.3%. The intermediate-layer latent space, in other words, is easier to search whether or not you have a good compass.

The honest caveats are the obvious ones. The evaluations run on smaller open backbones (LLaMA-3.2-3B through Qwen2.5-14B and Qwen3-4B) on math and graduate-level science QA, and the paper does not report per-query compute cost, memory overhead, or behaviour on much larger or already reasoning-tuned models. The authors also flag that concurrent Anthropic work called J-lens explores a related intermediate-layer idea but frames it for explainability rather than reasoning gains.

The forward-looking part is who this helps. Teams that cannot fine-tune the frontier get a plug-in inference-time lever that works across LLaMA and Qwen families, and the token-attribution result, that reasoning connectors like 'because' and 'therefore' absorb the strongest gradients, is a real interpretability handle. If those numbers replicate on bigger models, the case for spending compute at inference rather than in another training run gets a little stronger.