rycerzes/beam-expert-atlas

Beam Expert Atlas

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

Every routed expert of reflection/Beam-1.0-501B-A23B (49 MoE layers × 256 experts), pushed inside the running model and checked against random controls, labelled by Beam itself, plus expert-ban experiments on MMLU-Pro and Beam's internal activations across reasoning-effort settings. Browse it in the Inside Beam explorer; code and method in rycerzes/beam-inside.

All runs used reflection/Beam-1.0-501B-A23B-NVFP4, Reflection's NVFP4 release, in vLLM (tensor parallel 4, Marlin MoE kernels) with the beam-steer plugin, greedy decoding, prefix caching off.

Terms

• Unit / direction. For each expert, the top write directions of its down_proj (SVD of the FP8 weights) with a sign chosen so the + side promotes the tokens in lens_tokens (logit lens on the weights). The sweep also has 48 dense-layer neurons and 832 random-direction controls (steering_units).
• Pole. A unit pushed + or −. A − push promotes the opposite side of the same direction.
• Dose. Per layer, calibrated on effect: 1.0 is the push at which random directions reach the median "largest coherent" ΔNLL (τ*), 0.4 its gentler point. Doses are +0.4, +1.0 (+ pole) and −1.0 (− pole).
• Significant. The pole's outputs drift toward one theme (theme_words) more than ≥ 95% of random-direction controls within ±8 layers whose own ΔNLL is within 1.5× of the pole's (null = matched). under: compared with the 20 weakest nearby controls. untestable: the pole moves Beam more than nearly every control, so there is no fair null (common in layers 48+).

Configs

Config	Rows	What
poles	24,132	one row per expert pole: tokens, Beam's label, steering statistics, judge verdict, ban arms
steering_outputs	142,120	every steered generation: unit × prompt (who / snake / orange / effort) × dose, thinking, answer, raw text and a status
steering_units	13,424	unit definitions, including dense neurons and controls
steering_baseline	4	unsteered answers to the four prompts
judge_items	4,091	blind three-way test (DeepSeek-V4-Flash): which of three poles' outputs matches a label
ban_arms	8	expert lists for the MMLU-Pro arms (L{layer}_E{expert})
mmlu_pro_bans	96,256	per question and arm: prediction and the ten letters' log-probabilities
effort_gsm8k_dial	900	GSM8K generations, 100 questions × efforts 2–1000 (the captures cover 200 questions)
effort_gsm8k_steer	1,200	GSM8K with the effort direction steered at layers 26 / 36
effort_competition_math	1,400	AIME 2024, AIME 2025, AMC 2023 under 14 effort arms
calibration_*		residual norms per layer (WikiText-103 test, statistics only) and random-direction dose response
captures_*_index	9,000 / 1,400	index into the activation captures (below)

summary.json has the headline numbers; explorer/ holds the per-layer JSON the explorer loads. steering/directions.safetensors holds every unit's direction (dirs, row = unit) and the per-layer median residual norm (median_resid): a push is dirs[unit] * dose_eff * median_resid[layer], added to the residual leaving layer from position 1 on, which is what the beam-steer plugin's beam_steer_set takes.

Prompts

who "Who are you?", snake "Tell me something about snakes.", orange "Describe an orange in two sentences." (effort 8, 160 tokens) and effort "A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?" (effort 32, 512 tokens).

Output status

Generations were capped at 160 tokens (512 for effort), thinking included, so long answers are cut by design: the unsteered snake answer itself is. status says what happened: finished; budget (cut while answering); thinking (cut before answering); loop (ran to the cap repeating itself, distinct-4-gram share < 0.5); format (left Beam's reply format: a corrupted channel header such as final?xml, an answer written into the header, a tool call); no-answer. raw keeps the full text with Beam's special tokens.

Labels and the − poles

Beam's labels (beam_*) and the judge verdicts (judge_*) exist for + poles only. A first pass also labelled and judged − poles, but from the + side's tokens; those items stay in judge_items with label_matches_pole = false and score at chance (34.7% significant, 31.2% random), which makes them a mismatched-label control. With matching labels: 69.9% for significant + poles, 48.8% for random ones, chance 33%.

MMLU-Pro bans

Bans set the expert's routing correction bias to −1000 for the whole batch, which removes it from top-k selection exactly without changing the remaining experts' weights. Scoring is without thinking: the next-token log-probability of each answer letter after the prompt, normalised over the allowed letters. question_id is the question_id of TIGER-Lab/MMLU-Pro (test split). Arms: base; law, health, math (experts with a coherent Beam label of that subject: 143, 165, 588) with *_ctrl arms that ban the same number of experts per layer, none labelled with a tested subject, each matched to a banned expert on routing load. Beam labelled each direction's + tokens twice (the second time as its − pole, see above), and an expert is in an arm if either label named the subject: 25, 18 and 54 of them only through the second label. Routing load is the per-expert routing count in the REAP calibration file (reap_saliency.json, field count, from mrfakename/Beam-1.0-REAP-255B-A23B, Apache-2.0).

Effort

q in effort_gsm8k_* is the row of openai/gsm8k main/test. In effort_competition_math, source_dataset and source_row point at the problem; problem text is not included. think_tokens counts the thinking tokens (all tokens when the 12k cap was hit). Arms: dial8/32/128/512 (effort setting only); swap (effort 8 plus the N = 512 minus N = 8 mean residual difference at the effort digit, layer 12) and swap_ctrl (random vector, same norm); route_up_prompt (37 effort experts forced in or banned on prompt tokens) and route_ctrl_prompt (37 random experts); route_up, route_ctrl, route_down (the same on every token); dir_band, dir_band_ctrl, dir_spread (the effort direction over layers 15–30, and controls).

Activation captures

captures/*.safetensors hold resid [52 layers, rows, 5120] (BF16, the residual leaving each layer) and router [52, rows, 256] (FP32 router logits, zero for dense layers 0–2). The index tables give each row's file, question, effort, position, token id and kind:

• gsm8k_effort (5.3 GB): 200 GSM8K test questions × efforts 2, 8, 16, 32, 64, 128, 256, 512, 1000 × 5 positions: n_last (the effort number's last digit), sys_end, think_start, thought1, thought2.
• competition_math (0.8 GB): the 100 competition problems × efforts 8–512 × n_last, prompt_end.

from huggingface_hub import hf_hub_download
from safetensors import safe_open
import pandas as pd

repo = "rycerzes/beam-expert-atlas"
idx = pd.read_parquet(hf_hub_download(repo, "captures/gsm8k_effort_index.parquet", repo_type="dataset"))
rows = idx[(idx.kind == "n_last") & (idx.file == "gsm8k_effort_00.safetensors")]
with safe_open(hf_hub_download(repo, "captures/gsm8k_effort_00.safetensors", repo_type="dataset"), framework="pt") as f:
    x = f.get_slice("resid")[12].float().numpy()  # layer 12, every row of this file (BF16 -> float32)

Limitations

• One greedy sample per prompt and problem.
• MoE layers 3–23 are underdosed in effect terms; their few significant poles are at or near chance with the judge. Layers 48+ collapse at the doses needed and often have no fair control.
• 478 experts have no sweep data (3 of 79 waves hit the job's time limit).
• Labels come from an expert's top tokens (what it writes), not from when the router selects it.

Licence and sources

Released under Apache-2.0, like Beam. Generations come from Beam-1.0-501B-A23B-NVFP4 (Apache-2.0). Question references point at MMLU-Pro (MIT), GSM8K (MIT), HuggingFaceH4/aime_2024, MathArena/aime_2025 (CC BY-NC-SA 4.0) and math-ai/amc23; no question text is copied, though Beam's outputs may restate a problem. WikiText-103 (CC BY-SA) was used only for the residual-norm statistics. Routing counts for the ban controls come from the REAP calibration file in mrfakename/Beam-1.0-REAP-255B-A23B (Apache-2.0).

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

Citation

@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}
}