TL;DR

My main question is "Can WDD reveal useful information in native weight-transcribed writes that SOTA activation-based and learned-feature methods do not?"

My simple plan is to test this on open-source models using pre-specified frozen candidate lists, doing causal interventions as well as direct comparisons against today's circuit-tracing methods.

My minimum ask is $2.5k, to establish if WDD's native-write geometry identifies functional interference and suppression.

My max goal is $9.5k extending this to circuit-tracing, pre-cancellation info, learned-dictionary blindspots and replication on >7B models.

I will have succeeded if I find a reproducible capability which existing methods do not provide.

I will have failed if I show that WDD adds little to nothing beyond existing methods.

My output will be code, candidate lists, experiment logs, results, and a summary map of where WDD works and does not.

Project summary

I recently developed Weight-Dictionary Decomposition (WDD), which is a method for decomposing a transformer's residual-stream computation into write directions derived from the model's own weights. WDD is fairly ambitious, in the sense that it does not train an SAE, transcoder, probe, or any other representation model, thereby not relying on any learned feature dictionary at all. My accompanying paper and open-source repository test the method across eight transformer models (up to 7B parameters), spanning provenance and write-recovery tests, causal experiments, erasure/counter-writing analyses, as well as GPT-2 analysis.

My findings were mixed, as found in the paper. In summary, an identified write has the right sign in over 99.7% of cases in all 8 models. The write's magnitude is recovered with a median relative error 0.25-0.54 in seven of them, and identification rises with scale to 0.98 in Llama-7B. However, identification is capped by an observability ceiling in all the models. I find that a trained SAE reconstructs GPT-2 states better at matched sparsity compared to WDD. Further, the selected support is mostly not real writes in several models, and in pre-specified comparisons WDD beats a marginal outlier screen at nominating live circuits yet does not outperform a correlation screen.

Now, I want to acquire funding to systematically map where WDD is truly useful relative to the mech interp landscape. Namely, this project is entirely exploratory, investigating the potential of WDD without any claims being made towards solving a major interpretability problem.

Through my literature review, I found recent works have produced promising approaches including cross-layer transcoders and attribution graphs (Anthropic, 2025), natural language autoencoders (NLAs, Anthropic, 2026), the Jacobian lens and J-space (Anthropic, 2026), WeightLens and CircuitLens (Circuit Insights, ICLR 2026), parameter decomposition (APD and SPD), and Anthropic's analysis of interference weights. Against these, I aim to resolve a narrower claim against the above approaches:

What information about computation is exposed by the model's native write geometry that is difficult to obtain from activation-derived or learned feature representations?

I will answer it empirically along two directions, with the addition of two more if funding reaches the goal. These directions are as follows:

1. whether WDD can identify functional interference and counter-writing directly from native writes, and whether opposing native writes correspond to semantically meaningful suppression rather than simply geometric cancellation

2. whether WDD can provide useful candidate mechanisms or priors for existing activation-based circuit-tracing methods

3. (goal funding) whether WDD can identify information that is cancelled or transformed before it is visible in the residual stream

4. (goal funding) whether WDD finds mechanisms that are absent, fragmented, or difficult to identify in learned feature dictionaries

I welcome producing a negative result. More importantly, if WDD does not add usable information beyond SOTA methods, I strive to state exactly that instead of forcing a positive interpretation.

What are this project's goals? How will you achieve them?

I have several goals for this project. The first problem I aim to resolve is establishing a detailed comparison between WDD and current mechanistic-interpretability approaches. This is because the field has rapidly evolved since I started my WDD project.

In brief, existing literature such as cross-layer transcoders and attribution graphs provide increasingly powerful activation-based circuit tracing, while J-lens provides a way of reading semantically meaningful intermediate representations. Additionally, WeightLens and CircuitLens investigate learned weights and feature interactions, and recent Anthropic work characterizes interference weights, weight-implied links that need not be functional.

Based on the presence of these methods and more like it, I foresee WDD as a possible complementary hypothesis generator, yet also want to test whether WDD actually contributes anything unique.

Therefore, I will start with small open models and compare WDD-derived hypotheses against existing baselines wherever it is practical to do so. My experiments will focus on measurable questions instead of general claims which merely state that one interpretability method is superior.

In terms of achieving my goals, I have several main directions to pursue as listed above. Firstly, I want to track interference. Since WDD exposes native writes that point in opposing directions, I want to investigate whether these relationships can predict functional interference, semantic suppression, or distributed competition. To do so, I first identify candidate mechanisms from a frozen checkpoint before subsequently testing with forward passes and causal interventions. My success criterion reuses my prior work's Tier 1/2 tests to check whether on a pre-specified list of nominees, frozen before any intervention, the geometric prediction beats the correlation screen from the paper at picking interventions that change the model's output. Still within this first direction, semantic suppression is also of key interest for this project. In my existing WDD work, I conducted counter-writing and erasure experiments. This included a GPT-2 case where a simple eraser turned out to participate in a more distributed suppression/restoration mechanism. My research question here is: can native-write geometry identify what semantic representation is being suppressed, instead of just detecting that two writes oppose one another?

Secondly, I am eager to see whether WDD is complementary to current circuit-tracing methods. More clearly, without having to replace J-lens, transcoders, or attribution graphs, I will test whether WDD can generate useful candidate mechanisms before expensive activation-level searches occur. A benchmark I propose for this subsection is whether WDD reduces the number of activation examples or candidate features needed to be examined before finding a causally relevant mechanism. Success is determined via measurement runs on the same prompts and the same small model, WDD-first against baseline-first, with the open circuit-tracing tools as the baseline.

Lastly, I am curious about whether WDD exposes native-write contributions that are difficult to view after cancellation. As a write can be deemed computationally important even if further computation cancels its visible contribution, WDD allows access to the native write before said cancellation, marking this as a natural test of the method. This direction, and the fourth, run only at goal funding.

My timeline is dedicating the first month to the first direction on GPT-2 and sub-1B models, with the nominee lists frozen before any intervention. Then on the second month I spearhead direction 2 against a circuit-tracing baseline on one small model, while month 3 goes towards replications (7B models only at goal funding), write-up, and release, with a progress update on this page at the end of each month.

All in all, the expected output is a reproducible map of where WDD helps, where it overlaps with existing methods, and where it fails. I will release code, experiment logs, candidate records, and negative results.

How will this funding be used?

I am seeking $2,500 minimum funding, with a goal of $9,500.

My initial experiments are deliberately cheap, and I already possess my existing WDD implementation + infrastructure used for published experiments. The funding is mainly for compute and research time to perform the new comparative experiments detailed above. I do not aim to reproduce the existing paper.

As for how the funding will be used, across a span of three months, my budget is as follows

• $1,200 (minimum) to $7,200 (goal) - dedicated research/engineering time, namely, implementing the comparative pipeline, automating candidate discovery and evaluation, analyzing results, and preparing the final public release. The minimum pays for four weeks at 10 hours a week, enough for the first direction on models under 1B parameters, while the goal pays for twelve weeks at 20 hours a week.

• $500 (minimum) to $1,200 (goal) - GPU compute, namely, an A100 rented via vast.ai/runpod.io at an expected cost of roughly $1/hour. For reference, my WDD paper's full set of experiments used around 50 A100-hours, so the minimum covers about ten times that. This will cover WDD construction, candidate searches, forward-pass experiments, as well as causal intervention sweeps. Whenever it is sufficient, I will use smaller/cheaper models.

• $200 (minimum) to $500 (goal) - model/API access, namely, one provider. These will be used for experiment development, automated analysis, interpretation of candidate mechanisms, and comparison against existing interp results.

• $100 - storage and experiment infrastructure, namely, model artifacts, intermediate dictionaries, per-token records, logs, and backups.

• $500 - contingency, namely, additional replication runs, unexpected compute costs, or small amounts of additional infrastructure.

Funding between the minimum and the goal unlocks weeks in the following order: the first direction to start, then #2 against a circuit-tracing baseline, then #3 and #4 as well as replication across the 7B models at the goal, with no change to the above-established research plan.

In addition, I will prioritize experiments by information value, where my order of focus is small models first, replication prior to scaling, and larger checkpoints only when a result passes initial tests. Without funding, the first direction runs on GPT-2 on my own machine over a longer period, and nothing above 1B parameters can happen.

Who is on your team? What's your track record on similar projects?

I am working independently on this project. My WDD paper's experiments were designed, run and audited with Claude under my direction, the text was drafted with it, and every number was checked against the result files by both of us, with this project being run the same way.

As for my background, I am a 2025 UofT CS graduate conducting research pertaining to LLMs in my free time while applying for work. My previous project was titled "Weight-Dictionary Decomposition: Reading a Transformer's Residual Stream with Its Own Writes". In this project, I evaluated WDD across eight transformer models (GPT-2 small, SmolLM2-135M, Qwen2.5-0.5B, Pythia-410M, OLMo-1B, Qwen2.5-7B, Pythia-6.9B, and Llama-7B).

My released repository includes the experimental scripts, result files, logs, and per-token records underlying the paper. My 30-page WDD writeup itself includes reconstruction versus attribution, computational provenance, the observability ceiling, generality across eight models on two corpora, write roles with null controls, the channel-447 circuit in GPT-2, and pre-specified nomination comparisons, with the audit tables as appendices.

Thus, this project is a continuation of a previously implemented and experimentally tested research pipeline with no need to build a method from scratch.

Repository: https://github.com/sunmoonron/weight-dictionary-decomposition

Paper: https://doi.org/10.5281/zenodo.22749478

What are the most likely causes and outcomes if this project fails?

The most likely failure is that WDD's native-write geometry does not contain enough semantic information to identify useful mechanisms beyond what is already obtained by current activation-based or learned-feature methods. I see many reasons why this could be true. The first is that native writes could be too distributed or polysemantic to correspond cleanly to semantic concepts. The second is that opposing writes may frequently reflect ordinary transformations rather than suppression. The third is that despite our apparent advantage of seeing pre-cancellation writes, it may not translate into better mechanistic predictions at all. Finally, current methods (the Jacobian lens, cross-layer transcoders, attribution graphs, weight-based feature analysis) might recover the relevant information more effectively.

With that being said, I believe a negative result would still prove useful. Having such a result would establish an empirical boundary on how much semantic and causal structure can be recovered from native parameter geometry alone. Furthermore, it would identify cases where activation-derived representations are truly necessary.

I aim to treat above expectations as a real possibility and will not define success as finding an impressive-looking circuit.

How much money have you raised in the last 12 months, and from where?

I have raised $0 in research funding over the past 12 months. My entire existing WDD project was done using free/low-cost compute dedicating my personal research time.