About

By Swapnil Dutta (swappy · rycerzes on Hugging Face · GitHub).

Inside Beam is a map of the 12,544 routed experts of Reflection's Beam-1.0-501B-A23B: 12,066 of them were pushed inside the running model and checked against random controls (478 lost their data to a time limit). The write-up is on rycerz.es, and the research log has every step.

How Beam was used

• The subject. Every push, re-route and ban ran inside Beam-1.0-501B-A23B-NVFP4, Reflection's NVFP4 release, in vLLM on 4× RTX PRO 6000 (HF Jobs), through our beam-steer plugin.
• The labeller. Beam named 10,926 of its own expert directions (name, description, subject, coherent or not) through HF Inference Providers; those labels chose the experts for the bans.
• The measuring stick. Its identity line, its effort setting and its benchmark answers are the outcomes.

What you are looking at

• Push. Each expert's main write direction was added to the residual stream at three doses (gentle, full, and full the other way), on four prompts, greedy, up to 160 tokens (512 for the effort prompt). The studio shows those real outputs; nothing runs live. Long answers are cut by the token budget by design, and each answer says how it ended: finished, cut by the budget, stuck in a loop, or a broken reply format.
• Significant means the outputs drift toward one theme more than 95% of random directions that move Beam by a similar amount. Verified means a blind judge also picked the expert's output from three given its top tokens, with high confidence (its top confidence score).
• Judges. The blind test was scored by DeepSeek-V4-Flash (deepseek-ai/DeepSeek-V4-Flash-0731, via HF Inference Providers on DeepInfra, no thinking); it sees a label and three steered outputs, never which one is the expert's. Pilots on 200 items used Claude Sonnet 5.5, Qwen3.8-Flash-Next and Beam itself as judges, to choose the judge.
• Effort dial. Router choices recomputed from captured router logits for 200 GSM8K questions at each effort setting, with Beam's own grouped top-8 rule.
• Ban. MMLU-Pro (12,032 questions, no thinking) with the experts Beam labelled with a subject removed exactly via the routing bias, compared question by question with a ban of the same number of random experts matched on routing load (per-expert routing counts from the REAP calibration file, reap_saliency.json in mrfakename/Beam-1.0-REAP-255B-A23B, Apache-2.0).

Limits

• One greedy sample per prompt; 100 competition problems for the effort arms.
• Early MoE layers (3–23) are underdosed and their few hits are at or near chance with the judge (27–42% vs 33%); layers 48+ collapse into repetition at the doses needed and often have no fair control.
• 478 experts have no sweep data (3 of 79 waves hit a time limit).
• Labels describe what an expert writes, read from its top tokens, not when the router picks it.

Credits

Where this started. Simon V's post quoting @_lyraaaa_'s two-neuron "evil Qwen" steering recipe, and Goodfire's essay Neural Networks Are Already Clean, which argues most units have one readable job and that you check a label by steering it.

Method.

• Lyra's neuron-explorer: steering-first labelling with random controls, doses in units of the median residual norm, the who / snake / orange prompts, and the research-log format.
• RhizoNymph/vllm, the steering fork of vLLM that neuron-explorer runs on: our beam-steer plugin's per-request steering table follows its design (the fork has no Beam support, so we wrote our own).
• Expert steering and bans: SteerMoE (arXiv 2509.09660) and MoTE (arXiv 2502.11096).
• The blind three-way label test: Gur-Arieh et al. (arXiv 2501.08319). A model describing its own units: Li et al. (arXiv 2511.08579). Reasoning strength read from activations: Sheng et al. (arXiv 2506.08390). Expert-level MoE interpretability: Herbst et al. (arXiv 2604.02178). The logit lens: nostalgebraist (2020).
• Static weight analysis of Beam: Barneyjm/beam-teardown.

Model, serving and compute. Reflection's Beam-1.0-501B-A23B and its NVFP4 release (Apache-2.0), served with vLLM following its Beam recipe; transformers' Beam implementation for the reference checks; HF Jobs for the GPU runs; HF Inference Providers for Beam's labels (Baseten) and the judge, DeepSeek-V4-Flash (DeepInfra). Pilot judges: Claude Sonnet 5.5, Qwen3.8-Flash-Next and Beam itself.

Data. MMLU-Pro, GSM8K, AIME 2024, AIME 2025, AMC 2023, WikiText-103 (statistics only). The ban controls are matched on per-expert routing counts from the REAP calibration file (reap_saliency.json in mrfakename/Beam-1.0-REAP-255B-A23B, Apache-2.0).

The site. Built on Thibaud Frere's research article template (CC BY 4.0). The push panel's strength control follows Neuron Studio, the recipes follow Neuronary, the map follows Neuronpedia's feature maps, and the route view follows the data-flow overview of Polo Club's Transformer Explainer. Charts with D3; the map layout with UMAP.

Tooling. Built with the help of Claude Code (Claude Opus 5.5).

The contest. Reflection and Hugging Face's Build with Beam (announcement). Not affiliated with Reflection.

References

• Cho, A., Kim, G. C., Karpekov, A., et al. (2024). Transformer Explainer: Learning LLM Transformers with Interactive Visual Explanation and Experimentation. arXiv:2408.04619.
• Cobbe, K., Kosaraju, V., Bavarian, M., et al. (2021). Training Verifiers to Solve Math Word Problems (GSM8K). arXiv:2110.14168.
• Dahlke, R., Klagges, H., Zecha, D., Merkel, B., et al. (2025). Mixture of Tunable Experts: Behavior Modification of DeepSeek-R1 at Inference Time. arXiv:2502.11096.
• Fayyaz, M., Modarressi, A., Deilamsalehy, H., Dernoncourt, F., et al. (2025). Steering MoE LLMs via Expert (De)Activation. arXiv:2509.09660.
• Frere, T. (2025). Research Article Template. Hugging Face Space.
• Goodfire (2026). Neural Networks Are Already Clean. already-clean.goodfire.com.
• Gur-Arieh, Y., Mayan, R., Agassy, C., et al. (2025). Enhancing Automated Interpretability with Output-Centric Feature Descriptions. arXiv:2501.08319.
• Herbst, J., Wermter, S., Lee, J. H. (2026). The Expert Strikes Back: Interpreting Mixture-of-Experts Language Models at Expert Level. arXiv:2604.02178.
• Kwon, W., Li, Z., Zhuang, S., et al. (2023). Efficient Memory Management for Large Language Model Serving with PagedAttention (vLLM). arXiv:2309.06180.
• Li, B. Z., Guo, Z. C., Huang, V., et al. (2025). Training Language Models to Explain Their Own Computations. arXiv:2511.08579.
• Lyra (2026). neuron-explorer: steering-first neuron labelling. github.com/lyramakesmusic/neuron-explorer.
• McInnes, L., Healy, J., Melville, J. (2018). UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv:1802.03426.
• Merity, S., Xiong, C., Bradbury, J., et al. (2016). Pointer Sentinel Mixture Models (WikiText). arXiv:1609.07843.
• nostalgebraist (2020). Interpreting GPT: the logit lens. LessWrong.
• Reflection (2026). Beam-1.0-501B-A23B. Hugging Face model card.
• Reflection (2026). Beam-1.0-501B-A23B-NVFP4. Hugging Face model card.
• Sheng, L., Zhang, A., Wu, Z., et al. (2025). On Reasoning Strength Planning in Large Reasoning Models. arXiv:2506.08390.
• Wang, Y., Ma, X., Zhang, G., et al. (2024). MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark. arXiv:2406.01574.

Cite

@misc{dutta2026insidebeam,
  title        = {Inside Beam: steering, labelling and banning the experts of Beam-1.0-501B},
  author       = {Dutta, Swapnil},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/spaces/rycerzes/inside-beam}},
  note         = {Code: https://github.com/rycerzes/beam-inside; data: https://huggingface.co/datasets/rycerzes/beam-expert-atlas}
}