ITR Atlas
← All interfaces

ChatBCI-Assist: LLM-Assisted P300 Speller (Hong et al., 2026)

Measured by Hong, Rao, Wang & Najafizadeh · IEEE Trans. Biomed. Eng. 73(8) (2026)

EEG P300 LLM-assisted 2026

Inputs

The measured or assumed values behind the calculations, each with its source.

chars = 995
Correct characters delivered across all 30 Copy-LLM tasks (10 subjects × 3 sentences, 1,020 target characters). Scored from the public experiment logs as target length minus Levenshtein distance to the final text. 29 of 30 tasks finished exactly; one (S04, self-generated) timed out at 10 min with 54.5% character accuracy, matching Table II footnote e.
time = 97.5 min
Total wall-clock time for the same 30 tasks, first to last log timestamp (mean 3.25 min per task, matching Table II's 3.3 min T2C). The paper defines T2C as including LLM generation, system latency and user pauses (§II-D).
CPM_reported = 19.7 char/min
Authors' mean copy-spelling CPM (Table II). Not wall-clock: Eq. 7-8 compute it as characters per selection × 60 / T, with a nominal T = 3 s review + mean flashes × 0.14 s. That excludes the local LLM's 3.57 ± 1.94 s inference latency (§V) and other overhead. The logs show 14.6 s per selection in real time. Table II's own averages give 34 char / 3.3 min ≈ 10.3 char/min.
P_sel = 0.8846
Mean online selection accuracy over the 10 subjects (Table I: 88.9, 92.5, 90.1, 85.2, 84.6, 86.7, 92.0, 85.5, 91.2, 87.9%). Averaged over all nine online tasks per subject, i.e. all three sessions, not Copy-LLM alone.
H = 1.0 bits/char
English-text entropy (Shannon), the same ~1 bit/char standard applied to every character speller in the atlas. LLM word and phrase completion raises characters per selection, but it cannot raise the information content of English text above its entropy, so crediting delivered text at 1 bit/char already absorbs the language-model gain.
ITR_reported = 105.2 bits/min
Authors' mean copy-spelling ITR (Table II), Eq. 9: Wolpaw bits over k = 42 keys, with P = character accuracy, multiplied by CPM rather than selections/min. It therefore credits every character, including each character of an LLM-inserted phrase, with ~5.3 bits (105.2 / 19.7 = 5.34 ≈ log2 42). The authors acknowledge that the uniform-prior assumption does not fit an LLM speller (§V). The character-level MIR (52.9 bits/min) and semantic ITR (147.1 bits/min, semantic-spelling session) share the same nominal CPM and are not scored.

Strictest ITR

Each scoring method is an upper bound on the channel, so the headline is the strictest (smallest) one for this entry. Use the score selector on the home page to view any single method across entries.

Strictest Shannon (text) Recomputed
Character-entropy throughput (copy-spelling, wall-clock time)
Net of English redundancy and decoding error
0.17 bits/s

  1. Correct characters per minute

    995 correct char / 97.5 min = 10.2 correct char/min

    Pooled over all 30 Copy-LLM tasks from the public logs, including the one timed-out task. It cross-checks against Table II's averages: 34 char × 98.5% character accuracy / 3.3 min T2C ≈ 10.1 char/min. The mean of per-task rates is higher (13.2 char/min) because short, fast sentences weigh equally with long ones. The pooled ratio is the sustained rate.

  2. Bits per character

    H(English) ≈ 1.0 bit/char (Shannon)

  3. Information transfer rate

    10.2 char/min × 1.0 bit/char ÷ 60 s/min = 0.17 bits/s

    About 2.5× the checkerboard P300 speller (Townsend 2010, 0.067 bits/s). The dictionary baseline in the same study, same GUI and adaptive stopping, delivers 6.1 correct char/min (0.10 bits/s) on the same logs, so the fine-tuned LLM accounts for ~1.7×.

What counts as a bit depends on the action space. The number of distinguishable actions and how likely each one is are design choices of the task, not the sensing hardware. The same modality can present a fixed set of targets, a set pruned per step by a grammar or language model, or a continuous control space. Each of these changes how many actions are live and how the probability mass is spread, and therefore the information per selection. Read the action space below before comparing headline numbers across entries.

Action space

What the user can produce at each step, and how those options are distributed.

Structure
Context-dependent (the live set changes per step)
Size
42 distinguishable actions
Prior
Context-conditioned: likelihoods depend on prior actions
Notes
Row-column P300 speller on a 6 × 7 grid (§II-B): 26 letters, 4 function keys (delete word, delete character, space, enter), and 6 word ('+') plus 6 phrase ('*') keys that are filled each step by a LoRA-tuned Llama 3.1-8B running locally, trained on an ALS message bank. Bayesian adaptive stopping ends flashing once one key reaches 0.90 posterior (max 104 flashes); mean 51 flashes per selection. The 12 suggestion keys change content every selection, so the action space is context-dependent and Wolpaw's uniform prior does not hold. Ten healthy participants (12 recruited, 2 failed calibration); the ALS framing is the target population, not the tested one. The ranked figure is the copy-spelling session (verbatim target), with wall-clock time. The semantic-spelling session (30.7 CPM, users paraphrase the target) is not scored, because its output is not checked character-for-character. The authors' 19.7 CPM uses a nominal per-selection time that omits LLM latency, so it is kept only as a supplementary figure. Wall time also caps this entry against older spellers whose rates were back-derived from nominal timing (e.g. Townsend 2010).

Comparability The strictest bound here is the Shannon entropy of the output text, under one predictor held constant across the whole atlas (≈1 bit per character). That shared predictor makes it directly comparable to every other text entry (keyboards, spellers, silent speech and speech BCIs) regardless of prior or vocabulary size. For most text interfaces it comes out tighter than the raw-selection bounds, but not always. Where a small vocabulary makes Wolpaw tighter, that wins instead. Any Fitts, Wolpaw or log₂(N) figure shown below is another bound on the same channel. Switch the home-page score selector to compare one across entries.

Other score types

Bounds the atlas keeps out of the default strictest headline: as-reported figures, alternate task conditions, or raw-channel ceilings that shouldn't win the headline by default. Each still carries a score type, so the home-page selector ranks this entry on it when you choose that type. Read its derivation before comparing across entries.

Shannon (text) Author-reported · reproduced
Character-entropy throughput at the authors' nominal CPM
Authors' timing model, excludes LLM latency; shown for comparison
0.33 bits/s

  1. Characters per minute (authors' definition)

    CPM = (# characters / # selections) × 60 / T, with T = 3 s + f̄ × 0.14 s (Eq. 7-8) → mean 19.7 char/min (Table II)

    T counts only the 3 s suggestion-review window and the flashes. Wall-clock time per selection in the logs is 14.6 s; the gap is mostly the ~3.6 s local LLM inference plus system and user pauses.

  2. Information transfer rate

    19.7 char/min × 1.0 bit/char ÷ 60 s/min = 0.33 bits/s
Wolpaw Author-reported · reproduced
Authors' ITR: Wolpaw over k = 42 keys, scaled by characters/min
Per-selection classifier bits credited per character, shown for comparison
1.71 bits/s

  1. Bits per character (authors' Eq. 9)

    B = log2(42) + P·log2(P) + (1 − P)·log2((1 − P)/41), P = character accuracy. At P = 0.985: B ≈ 5.20 bits

    Eq. 9 feeds final character accuracy (98.5% mean) into a formula meant for per-selection accuracy, then multiplies by characters rather than selections per minute.

  2. Information transfer rate

    ITR = 5.20 bits/char × 19.7 char/min = 102.4 bits/min ÷ 60 = 1.71 bits/s

    Cross-check: the authors report 105.2 bits/min (1.75 bits/s, Table II), the mean of per-task B × CPM. The ratio 105.2 / 19.7 = 5.34 bits/char sits between B at the mean accuracy and log2 42 = 5.39 because most tasks finished at 100% accuracy. Either way it credits each LLM-inserted character with ~5.3 bits and uses the nominal CPM, overstating both the per-character information and the rate.

Wolpaw Recomputed
Wolpaw bitrate over N = 42 keys per selection (wall-clock)
Uniform 1-of-42 key-selection channel, shown for comparison
0.293 bits/s

  1. Bits per selection (Wolpaw formula)

    B = log2(N) + P*log2(P) + (1-P)*log2((1-P)/(N-1))
  = log2(42) + 0.8846*log2(0.8846) + 0.1154*log2(0.1154/41)
  = 4.258 bits / selection

    Term 1 is the information if every choice were correct; terms 2-3 subtract the bits lost to the error rate, assumed spread evenly over the other N-1 targets.

  2. Selections per second

    T = 14.55 s/selection  ->  1 / 14.55 = 0.069 selections/s

  3. Information transfer rate

    ITR = B * selections/s = 4.258 * 0.069 = 0.293 bits/s
Nuyujukian Recomputed
Nuyujukian achieved bitrate over N = 42 keys
Achieved-bitrate view of the key-selection channel, shown for comparison
0.28 bits/s

  1. Achieved-bitrate credit per net-correct selection

    N = 42 → log2(N − 1) = log2(41) = 5.358 bits per net-correct selection (field-standard achieved bitrate).

  2. Net-correct selection rate

    net-correct = 2P − 1 = 2(0.8846) − 1 = 0.769 of selections. At 1 / 14.55 s (402 Copy-LLM selections in 97.5 min) → 0.769 / 14.55 = 0.0529 correct/s.

    Selection accuracy is Table I's all-session mean; Copy-LLM alone is not broken out. A wrong key commits an error that must be deleted, so incorrect = 1 − P.

  3. Achieved bitrate

    5.358 bits × 0.0529 correct/s = 0.28 bits/s

Source

Authors
Hong, Rao, Wang & Najafizadeh
Publication
IEEE Trans. Biomed. Eng. 73(8), 2026
Paper
10.1109/TBME.2026.3693965
Reference
Hong et al. 2026 (IEEE TBME): 'ChatBCI-Assist: An Intent-Based P300 Speller With A Locally Deployed LLM and Adaptive Stopping Strategy Enabling Record Online Spelling Performance'
Reference
ChatBCI-Assist experiment logs (per-selection timestamps and text for all 90 online tasks)