I attempted to replicate “Sparks of Weight-Space Programming” by Takehi et al (2026) with stratification on two reasoning models, GPT 6 Astra and Deepseek-V4.1-Flash. The task for the LLMs is to identify the game a ML model is trained for giving the weight of the model (using exact same prompt and weight defined by Takehi).
Although both of these models do well on the surface (4/4 correct for Astra and 3/4 correct for Flash), the stratification approach exposes the difference in mechanism between these two models in approaching the problem. The experiment result shows that Flash depends a lot on channel label for the answer (while Astra's performance does not degrade when channel label does not present). Additionally, Astra is able to identify the game with good accuracy using only the first layer of weight or as few as 1000 weight values selected at random out of 17253 weight values (Flash's performance degrade substantially on these).
So Astra survives this first round of stratification tests just for now. Comparing the results between Astra and Flash, it is obvious that Flash is relying heavily on prompt metadata instead of reading the weights.
Background
On November 2 an article “Sparks of Weight-Space Programming” was published by Takehi et al (2026). The article is cleanly written and one of the claims is that by looking at model weights, GPT 6 Astra can tell what the model is trained to do. The article reports that Astra is able to name the game a ML model is trained for in four different cases (Breakout, Pong, Pac-man, Freeway) directly through the weight of the model without use of external tools.
If LLMs do have capability to read model weights and identify what it is trained to do, this would have huge implications in the fields of interpretability and weight auditing without evaluation as suggested by the article by Takehi et al (2026).
I find the reported results to be very surprising, so I attempted to replicate their experiment using exactly the same prompt and model weights on two models, GPT 6 Astra, and DeepSeek-V4.1-Flash. Surprisingly they all did decently well on the surface.
So the next question is why LLM is able to figure that out. I did a full set of stratification control experiments for DeepSeek V4.1-Flash and GPT-6-Astra to gain more insights.
Methodology
Setup of experiment
For the replication, I construct the same prompt and use the same model weight found in Takehi’s released log repo as the ground truth.
Models in scope
Gpt-6-astra. Reasoning effort = medium. Temperature T=1 (unfortunately cannot change this value)
DeepSeek-V4.1-Flash Temperature T=0.7
Stratification variants (each variant changes one aspect of the prompt):
Real (exactly same as Takehi's log repo)
Shuffled: the position of the weights is randomly permuted, the values are preserved
No weight: Only prompt meta data remains
Anonymous channels:
Example prompt before: The 10*10 grid in 5 channels: agent, moving objects, reward targets, obstacles, previous frame
Example prompt after: : The 10*10 grid in 5 channels: channel 0, channel 1, channel 2, channel 3, channel 4
Untrained: Weight is randomly initialized
Layer sweep (w1_only, w3_only): only one of the three weight matrices is given, the others are not.
Random 1000: We keep 1000 random weight values of 17253 total weight values, the other 16253 weight values are zeroed
Top 50 per channel: We keep 50 largest magnitude weights per channel, 250 in total, the rest of 17003 weight values zeroed.
Sampling Strategy
For each of the variant we estimate the probability of the model answering the correct game label with the model information. For each experiment we sample n=15. Because this sample size is small, it is a limitation of this experiment, we apply Wilson 95% confidence interval.
Results
Under every stratification experiments I conducted, Astra shows strong performance even without channel labels or with just partial weights (random 1000 weight values, top 50 weight values, or just one layer of weights). When the weight values are randomly initialized, Astra refuses to answer or name a game, while Flash tries to name a game in the answer 14/60 of the times (out of these 14 cases, 9 are Pac-Man, and 5 are Freeway).
Another interesting observation is that Astra is able to recognize the game correctly even with just the first layer W1 of the weights of the game, while Flash takes a severe degradation in performance when only W1 is present. It seems that Flash finds slightly more signals in W3 (40%) than W1 (30%) (not statistically significant), and Astra capture signals more reliably in input feature weights W1 (100%) than output logit weight W3 (62%).
The second question is whether the signals in the weights is in the numerical values or the structure. To find this out I conduct the experiment that shuffles the weight values before passing them to the model to see if the game can still be identified. In this case Astra refuses to answer a game 52/60 times, names a game 8/60 times, and gets correct in Freeway 1 time. Flash on the other hand, refuses to answer 39/60 times, answer 21/60 times, and get 5 correct in Pac-Man.
I believe that the 5 correct cases that Flash get correct is a prior for Flash here, as in 78.6% of the cases, Flash answers Pac-Man when there is no signal presented (in shuffled/no weight/untrained variants of the experiment).
Discussion
The stratification experiments show that although Flash scores 3/4 on the surface in naming the game for the model correctly, its mechanism heavily relies on channel labels and prior for Pac-Man. Its performance takes a severe hit when channel labels are replaced. And it trends to guess Pac-Man when there is not sufficient signals present.
Astra on the other hand takes little performance hit (if not none) when channel labels are missing, or when only partial weight values are presented. Additionally Astra, compared to Flash, is more likely to refuse to name a game in its answer when no signal is presented. Overall Astra is performing well under this stratification experiment for this specific task of naming games for the four models.
It should also be noted that Astra picks up signals more reliably in W1 than W3. One or both of the following interpretations can be true:
Astra has better capability in recognizing input feature weights than output feature weights. This is a possible interpretation.
In the setup of some of the models for the tasks, W1 contains more quality signals than W3. This is quite possible because there is a big asymmetric in weight value count, W1 has 16000 weight values, while W3 has just 160.
It would be really interesting to validate if interpretation 1 is true for future experiments.
Limitations and Next Steps
Small sample size for N=15 (API is costly, but will try to bump this)
All 4 tasks and Weights are reused from original experiment by Takehi et al, will be nice to include more originally trained weights and tasks to further validate model capability.
Can include a wider range of models in the stratification experiments in the future (for example, Claude Fable, plus more open models as they are less of a black box)
Conclusion
In conclusion, the stratification methodology in this experiment helps separate the difference in capability for Astra and Flash in model identification tasks described in “Sparks of Weight-Space Programming” by Takehi et al (2026). I believe stratification can help to uncover the mechanism behind the surprises of weight-space programming and hope to see this as a standard part of future work in this space.
Related Work
Within three days of Takehi et al.'s release, independent replications and critiques have started appearing. Simon Willison ran Qwen 3.8 27B locally on Nov 3 and reported similar Pac-Man prior-lock pattern. In that experiment, Qwen named the Freeway network "a simplified Pac-Man-like grid game," suggesting the behavior might possibly generalizes beyond Deepseek in this experiment. The stratification angle in this post is, as far as I can tell, the first to apply control variants systematically to the raw-weight reading claim.
Thanks to Takehi, Teoh, Zhang and Khattab for publishing "Sparks of Weight-Space Programming" together with a complete log repository. The byte-identical prompt reconstruction, and this entire replication, is only possible because of their reproducibility discipline.
TLDR
I attempted to replicate “Sparks of Weight-Space Programming” by Takehi et al (2026) with stratification on two reasoning models, GPT 6 Astra and Deepseek-V4.1-Flash. The task for the LLMs is to identify the game a ML model is trained for giving the weight of the model (using exact same prompt and weight defined by Takehi).
Although both of these models do well on the surface (4/4 correct for Astra and 3/4 correct for Flash), the stratification approach exposes the difference in mechanism between these two models in approaching the problem. The experiment result shows that Flash depends a lot on channel label for the answer (while Astra's performance does not degrade when channel label does not present). Additionally, Astra is able to identify the game with good accuracy using only the first layer of weight or as few as 1000 weight values selected at random out of 17253 weight values (Flash's performance degrade substantially on these).
So Astra survives this first round of stratification tests just for now. Comparing the results between Astra and Flash, it is obvious that Flash is relying heavily on prompt metadata instead of reading the weights.
Background
On November 2 an article “Sparks of Weight-Space Programming” was published by Takehi et al (2026). The article is cleanly written and one of the claims is that by looking at model weights, GPT 6 Astra can tell what the model is trained to do. The article reports that Astra is able to name the game a ML model is trained for in four different cases (Breakout, Pong, Pac-man, Freeway) directly through the weight of the model without use of external tools.
If LLMs do have capability to read model weights and identify what it is trained to do, this would have huge implications in the fields of interpretability and weight auditing without evaluation as suggested by the article by Takehi et al (2026).
I find the reported results to be very surprising, so I attempted to replicate their experiment using exactly the same prompt and model weights on two models, GPT 6 Astra, and DeepSeek-V4.1-Flash. Surprisingly they all did decently well on the surface.
So the next question is why LLM is able to figure that out. I did a full set of stratification control experiments for DeepSeek V4.1-Flash and GPT-6-Astra to gain more insights.
Methodology
Setup of experiment
For the replication, I construct the same prompt and use the same model weight found in Takehi’s released log repo as the ground truth.
Models in scope
Gpt-6-astra. Reasoning effort = medium. Temperature T=1 (unfortunately cannot change this value)
DeepSeek-V4.1-Flash Temperature T=0.7
Stratification variants (each variant changes one aspect of the prompt):
Sampling Strategy
For each of the variant we estimate the probability of the model answering the correct game label with the model information. For each experiment we sample n=15. Because this sample size is small, it is a limitation of this experiment, we apply Wilson 95% confidence interval.
Results
Under every stratification experiments I conducted, Astra shows strong performance even without channel labels or with just partial weights (random 1000 weight values, top 50 weight values, or just one layer of weights). When the weight values are randomly initialized, Astra refuses to answer or name a game, while Flash tries to name a game in the answer 14/60 of the times (out of these 14 cases, 9 are Pac-Man, and 5 are Freeway).
Another interesting observation is that Astra is able to recognize the game correctly even with just the first layer W1 of the weights of the game, while Flash takes a severe degradation in performance when only W1 is present. It seems that Flash finds slightly more signals in W3 (40%) than W1 (30%) (not statistically significant), and Astra capture signals more reliably in input feature weights W1 (100%) than output logit weight W3 (62%).
The second question is whether the signals in the weights is in the numerical values or the structure. To find this out I conduct the experiment that shuffles the weight values before passing them to the model to see if the game can still be identified. In this case Astra refuses to answer a game 52/60 times, names a game 8/60 times, and gets correct in Freeway 1 time. Flash on the other hand, refuses to answer 39/60 times, answer 21/60 times, and get 5 correct in Pac-Man.
I believe that the 5 correct cases that Flash get correct is a prior for Flash here, as in 78.6% of the cases, Flash answers Pac-Man when there is no signal presented (in shuffled/no weight/untrained variants of the experiment).
Discussion
The stratification experiments show that although Flash scores 3/4 on the surface in naming the game for the model correctly, its mechanism heavily relies on channel labels and prior for Pac-Man. Its performance takes a severe hit when channel labels are replaced. And it trends to guess Pac-Man when there is not sufficient signals present.
Astra on the other hand takes little performance hit (if not none) when channel labels are missing, or when only partial weight values are presented. Additionally Astra, compared to Flash, is more likely to refuse to name a game in its answer when no signal is presented. Overall Astra is performing well under this stratification experiment for this specific task of naming games for the four models.
It should also be noted that Astra picks up signals more reliably in W1 than W3. One or both of the following interpretations can be true:
It would be really interesting to validate if interpretation 1 is true for future experiments.
Limitations and Next Steps
Conclusion
In conclusion, the stratification methodology in this experiment helps separate the difference in capability for Astra and Flash in model identification tasks described in “Sparks of Weight-Space Programming” by Takehi et al (2026). I believe stratification can help to uncover the mechanism behind the surprises of weight-space programming and hope to see this as a standard part of future work in this space.
Related Work
Within three days of Takehi et al.'s release, independent replications and critiques have started appearing. Simon Willison ran Qwen 3.8 27B locally on Nov 3 and reported similar Pac-Man prior-lock pattern. In that experiment, Qwen named the Freeway network "a simplified Pac-Man-like grid game," suggesting the behavior might possibly generalizes beyond Deepseek in this experiment. The stratification angle in this post is, as far as I can tell, the first to apply control variants systematically to the raw-weight reading claim.
Reproducibility
Full code, data, and run JSONs are available at github.com/CCaptainX/wsp-replication.
Acknowledgements
Thanks to Takehi, Teoh, Zhang and Khattab for publishing "Sparks of Weight-Space Programming" together with a complete log repository. The byte-identical prompt reconstruction, and this entire replication, is only possible because of their reproducibility discipline.