SamplerScope: Exact decoder attribution for language agents

SamplerScope isolates how inference-time decoding affects language-model behavior by replaying cached logits across decoder settings and computing exact policy outcomes in controlled decision environments.

Endorsed byYYonathan

Project Details

SamplerScope tests whether behavior attributed to language model weights can instead be caused by the inference-time decoder. It caches grammar-constrained action logits from two Qwen2.5-Instruct checkpoints in two finite decision environments, applies 11 decoder configurations to the same logits, and computes each induced policy's outcomes exactly using dynamic programming. Across 24 controlled model-environment-label strata, greedy decoding improved benchmark return in 12 and reduced it in 12; decoder-only changes ranged from -1.224 to +0.313. Top-p sometimes removed all benchmark-optimal actions across 18.5% to 73.3% of decision occupancy, while exhaustive A/B/C mappings exposed strong surface-label sensitivity. SamplerScope provides a reproducible decoder-attribution toolkit and shows that observed agent behavior belongs to a model-prompt-grammar-decoder system, not model weights alone. It measures operational policies, not intrinsic preferences or sentience.

Hackathon paper: SamplerScope: Exact decoder attribution for finite language-agent behavior
Code: github.com/unshDee/sampler-scope
Other: github.com/unshDee/sampler-scope/tree/main/results

Theory of Impact

A lot of behavioral evaluations of language models treat the action produced by the model as evidence about the model itself. SamplerScope showed that this can be misleading in some settings because the decoding method applied after the model produces its logits can change the resulting policy and downstream outcomes.

The original study was deliberately controlled, so the important next question is how much this matters outside those small environments. If the same sensitivity appears across stronger models, different model families, and more realistic action formats, it would suggest that decoding should be treated as an important experimental variable when interpreting model behavior.

The practical goal would be to not to argue that one decoder is better than another. It is to help researchers identify when a result is stable across reasonable decoding choices and when it is not. This could make behavioral and safety evaluations easier to interpret and reproduce. A negative result would also be useful, as it would establish where the effects found in the original study stop being important.

People

Ansh Dawda

Team Member

Track Record

Digital Minds Research Sprint (Apart Research), 2026

Grants Received– no grants recorded

Add grant

Funding Asks

SamplerScope: Exact decoder attribution for finite language-agent behavior

Open

Minimum

$10,000

Ideal

$20,000

What this ask is

I would use the funding to extend SamplerScope beyond the fairly controlled setting of the original project. The first study showed that changing only the decoder, while keeping the underlying model logits and environment fixed, could noticeably change the resulting agent behavior and sometimes even change the direction of the effect. However, it was intentionally kept small, I used two Qwen 2.5 checkpoints, two synthetic environments, and one-token A/B/C actions.

The first thing I want to test is whether the same effect is still present with larger models and models from unrelated families. This would directly address one of the main limitations of the original work. I would also extend the setup beyond single-token actions to semantic labels, multi-token actions and, if feasible, structured or tool-call actions.

Another part of the project woudl be to test a few additional sequential environments with different horizons, action spaces, and reward structures. I am particularly interested in whether properties such as logit margin or entropy can predict when an evaluation will be sensitive to the choice of decoder.

Finally, I want to look beyond changes in return and ask whether decoding can actually change the conclusion of an evaluation, for example by reversing the ranking between two models or changing an observed behavioral tendency. The outputs would include the extended experiments, reproducible traces and code, an improved SamplerScope evaluation toolkit, and a research report or preprint.

How the money will be spent

For the $10k minimum to the $20k ideal funding level, I would roughly allocate the budget as follows:

Research time: $5,000 - $9,000
GPU compute / model inference: $3,000 - $6,000
Infrastructure, experiment storage, hosting: ~$1,000
Additional replication / external methodological review: up to $2,500 (depending on funding)
Dissemination, contingency, unexpected experiment reruns: $1,000 - $1,500

Research time would cover experiment design, implementation, model and decoder integration, validation, analysis, reproducibility work, and writing the final research report or preprint.

At the minimum funding level, I'd prioritize the core experiments, compute, and research time. Additional funding would allow me to expand the model and environment coverage, run more extensive replication and robustness checks, and potentially involve external methodological review.

Discussion

Private comments in this thread are only shown to approved funders and grant reviewers. Email hi@grantmaking.ai or to request access.

Private UserPrivate15d?

AAnonymous judgeJudge · Digital Minds Research Sprint15d

This is actually a very good study, the traces and controls set high bar for reproducibility. The results are from qwen checkpoints with 2 environments and 42 states with 1 letter action labels. It shows weakness of small model on multiple choice prompts rather than decoders.

YYonathan ArbelJudge · Digital Minds Research SprintEndorsement15d

The paper shows, quite convincingly, that observed agent behavior can be highly dependent on the specific decoder used. This is taken to mean that research on model preferences should also control for the model's decoder. I agree with this point, but note that many safety evaluations use representative settings and evaluate the system-as-deployed. Still, the project is valuable and interesting and it exposes another "degree of freedom" in model assessment.