Such low dimensionality of the LoRA vector must surely result in a close-to-linear modification to the KV calculation. This seems to me to imply that what we call "reasoning" is latent within the model. Pretty clear I didn't read the paper, I'm sure the authors address this.
>One theory is that the knowledge required to solve the task is already stored in the parameters of the model, and only the style has to change for task success
>In particular, learning to generate longer outputs may be possible in few parameters
>we develop budget forcing to control test-time compute by forcefully terminating the model’s thinking process or lengthening it by appending “Wait” multiple times to the model’s generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps
Maybe, indeed, the model simply learns to insert the EOS token (or similar) later, and the capability is already in the base model
Not sure if I buy it. First, SVD decomposition to obtain U, Σ, V is computationally expensive, so it would work only if we are not finetuning very big models.
But my real concern comes at the results. The "13 parameters" looks like bait, because it is one result of finetuning a model on a very simple math benchmark, grade-school-math (GSM8K), an already very saturated benchmark on every model. Besides, it seems to happen only for the qwen family model... It looks like GSM8K was part of the training set of the qwen model, and this tinylora finetuning did the last adjustments to perfectly reflect that overtraining.
I've done a lot of exploratory work with Stable Diffusion LoRAs, and I actually do buy that there's some juice here, though it's almost certainly not nearly as good as other techniques can be. In particular, this technique will likely avoid the intruder dimension problem which plagues naive LoRA. SVD is expensive, but you only have to do it once at the beginning of training.
I haven't done much research lately, but when I was working on it, I was having substantial success training an adapter of the form U_k @ P @ A, where U_k was the top k left singular vectors of the underlying weight, and then P and A were your typical LoRA projection matrices.
The 13 parameters are kind of misleading here; the real juice is going to be in the P_i fixed random matrices. My suspicion is that they are overfitting to the benchmark, but they almost certainly are observing a real gain in model capacity that is largely due to avoiding the intruder dimension problem.
If 13 parameters can unlock better reasoning, then we will not be "training" models, we'll be steering them. Most of the capability is already there.
The real unlock isn’t TinyLoRA, it’s what this implies: ultra-cheap, continuous adaptation. The bottleneck shifts from compute to having a good reward signal.
Thats a wonderful explanation (and roughly the conclussion I arrived at after browsing the paper), I just wish it would have been in the original post.
This is interesting and all, but “LoRA” is painfully close to “LoRa” (which is related to radio networking, not AI) when just scanning a list of topics. We’re never going to beat the Shannon limit on acronyms and initialisms.
I’m glad the rest of the anchor text gave some context.
It's not "13 parameters to reason", they just rotated the full 8B parameter space in 13 dimensions and found a rotation that was still able to reason.
Depending on the latent structure, it's possible a nice rotation that would be perfect for some one specific problem, but you still got to search for it, and it's not a guarantee to exist.
But it's a nice step towards LLM parameter-space interpretability.
22 comments
[ 3.0 ms ] story [ 49.6 ms ] thread[0]: cartesien.io or Salesforce's WebscaleRL
>In particular, learning to generate longer outputs may be possible in few parameters
Reminded me of: https://arxiv.org/abs/2501.19393
>we develop budget forcing to control test-time compute by forcefully terminating the model’s thinking process or lengthening it by appending “Wait” multiple times to the model’s generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps
Maybe, indeed, the model simply learns to insert the EOS token (or similar) later, and the capability is already in the base model
But my real concern comes at the results. The "13 parameters" looks like bait, because it is one result of finetuning a model on a very simple math benchmark, grade-school-math (GSM8K), an already very saturated benchmark on every model. Besides, it seems to happen only for the qwen family model... It looks like GSM8K was part of the training set of the qwen model, and this tinylora finetuning did the last adjustments to perfectly reflect that overtraining.
I haven't done much research lately, but when I was working on it, I was having substantial success training an adapter of the form U_k @ P @ A, where U_k was the top k left singular vectors of the underlying weight, and then P and A were your typical LoRA projection matrices.
The 13 parameters are kind of misleading here; the real juice is going to be in the P_i fixed random matrices. My suspicion is that they are overfitting to the benchmark, but they almost certainly are observing a real gain in model capacity that is largely due to avoiding the intruder dimension problem.
The real unlock isn’t TinyLoRA, it’s what this implies: ultra-cheap, continuous adaptation. The bottleneck shifts from compute to having a good reward signal.
Let's say we have a low level programmer expert and we try to teach him algebra either we:
I’m glad the rest of the anchor text gave some context.
Depending on the latent structure, it's possible a nice rotation that would be perfect for some one specific problem, but you still got to search for it, and it's not a guarantee to exist.
But it's a nice step towards LLM parameter-space interpretability.