top-b/ Read paper

arXiv:2603.14567 · Preprint · March 2026

Let uncertainty
set the boundary.

Top-b is an adaptive decoding strategy that contracts around confident predictions and expands when the model is genuinely uncertain.

Explore the method View on arXiv
Deepon Halder AI4Bharat · IIEST Shibpur
Raj Dabre IIT Madras · AI4Bharat

Live probability field

t = 037

Normalized entropy 0.32
Adaptive bandwidth 0.33
Candidate set 2 / 16
top-b threshold
Confident / peaked Uncertain / flat

Drag uncertainty. Watch the viability band breathe.

01 / Core idea

The model's confidence governs its generative freedom.

Why it mattersStatic → Adaptive

Static cutoffs
do not listen.

Natural language shifts between sharp, near-deterministic states and broad, ambiguous ones. A fixed rule applies the same strictness to both.

k

Top-k

Counts tokens.

Always keeps a fixed number, regardless of the distribution's shape.

Fixed support
p

Top-p

Counts mass.

Keeps tokens until a fixed cumulative probability is reached.

Fixed threshold
b

Top-b

Measures uncertainty.

Anchors to the mode, then scales the admissible band with entropy.

Self-regulating
Low entropy Contract.

Prune distractors during reasoning.

High entropy Expand.

Preserve valid creative options.

The mechanism03 steps

A probability band
with a pulse.

Top-b replaces a global cutoff with a local, entropy-aware constraint. Every token step recalculates how far the sampler may move from the mode.

01

Measure uncertainty

H(p) = -Σ pi log pi

Shannon entropy tells us whether the next-token distribution is sharp or flat.

02

Scale the band

bt = b0 (1 + H(p) / Hmax)

Normalized entropy continuously expands the bandwidth from b0 toward 2b0.

03

Keep the viable set

Sb,t = { y : pt(y) ≥ (1 - bt)pmax }

Only tokens within the entropy-regulated relative band are renormalized and sampled.

Conceptual decoder

Five lines between logits and the next token.

Entropy is a vector operation over the vocabulary; the paper argues its cost is negligible relative to transformer attention. 

01 entropy  = -sum(p * log(p))
02 h_norm   = entropy / log(vocab_size)
03 band     = base_band * (1 + h_norm)
04 keep     = p >= max(p) * (1 - band)
05 next     = sample(normalize(p[keep]))
Empirical resultsGemma 3 4B IT

Less noise.
Stable reasoning.

On GPQA and preliminary GSM8K evaluation, Top-b reduces generation entropy and inter-run variance while retaining competitive reasoning accuracy.

GPQA average entropy

0.1579 Lowest among 7 tested strategies

GPQA accuracy

22.85% Competitive while constraining entropy

GSM8K accuracy

26.80% Highest among the 4 listed methods

GSM8K variance

1.136 × 10-4 About 9.9× lower than Top-p

GPQA / average sequence entropy

Sharper distributions, maintained accuracy.

Lower is better ↓

Top-b
0.1579
Eta
0.1592
Epsilon
0.1712
Min-p
0.1714
Top-p
0.1764
Top-k
0.1815

Accuracy and entropy are reported from the paper's GPQA evaluation. GSM8K uses a preliminary 1,000-sample Chain-of-Thought split.

Token-level casesAppendix D

Prune the tail,
not the signal.

In low-entropy contexts, cumulative mass can hide a surprisingly large candidate pool. Top-b keeps only tokens close enough to the mode.

Arithmetic2 + 2 =
Mode“4” · 39.75%
Top-b1token
Top-p14tokens
Syntactic continuation The battle that ensued…
Mode“was” · 46.00%
Top-b1token
Top-p44tokens
Open-ended instructionHelp me write an email…
Valid startsnewline · “I” · blank line
Top-b3tokens
Top-p43tokens
Moderate entropyDescribe a difficult decision.
Valid startsnewline · “You”
Top-b2tokens
Top-p25tokens
Original figuresFrom the paper

Inspect the
evidence.

Select a figure to explore the method's structure and measured behavior.

Four charts comparing Top-p and Top-b in low- and high-entropy states
01 · Static cutoff vs. adaptive band

Top-b contracts around the mode in low-entropy reasoning states and expands in high-entropy creative states.

Open full figure ↗
The paper11 pages

A self-regulating
language process.

Top-b frames autoregressive generation as a trajectory through a relative probability manifold. The sampler reacts online to local information density, pruning low-likelihood branches before drift accumulates.

Download PDF DOI

Cite this work

@article{halder2026topb,
  title   = {Top-b: Entropic Regulation of Relative
             Probability Bands in Autoregressive
             Language Processes},
  author  = {Halder, Deepon and Dabre, Raj},
  journal = {arXiv preprint arXiv:2603.14567},
  year    = {2026}
}

The principle

When the model knows,
make the path narrow.
When it wonders,
let the band breathe.
Read the full paper