NeurIPS 2026

SeismosA Statistical Signal Detection Framework for Semantic Chunking

Ishtiak Mahmud Saad¹
Mominul Islam²
Shahriyar Zaman Ridoy^3,6
Md Manjurul Ahsan^4,6
Azmine Toushik Wasi^5,6

¹Rajshahi University of Engineering & Technology
²Tulane University
³North South University
⁴University of Oklahoma
⁵University of Surrey
⁶Computational Intelligence and Operations Laboratory (CIOL)

Paper soon arXiv soon Code BibTeX

similarity signal · k = 2.5 x_i = cos(e_i, e_i+1)boundary, t_i > k

TL;DRSplit a document only where its sentence-similarity signal has a statistically significant valley. One parameter, no LLM calls, linear time.

5 / 5BEIR datasets where Seismos is the best chunker
+68–79%chunk nDCG@5 in end-to-end RAG
15–23×faster than production semantic chunkers
k = 2.5one parameter, never retuned

Abstract

Chunking is a hidden bottleneck in dense retrieval and retrieval-augmented generation: it determines the units that can be indexed, retrieved, and ultimately used as evidence. Yet most systems still segment documents using fixed token windows or globally thresholded semantic similarity drops, treating chunking as an engineering heuristic rather than a statistical decision. We propose Seismos, a statistical signal detection framework for semantic chunking. Seismos models consecutive sentence-embedding cosine similarities as a document-level signal and derives boundary decisions from the null hypothesis of no semantic transition.

Across four development BEIR corpora, we establish three corpus-invariant properties: bimodal document structure, rapid survival decay of shallow local minima, and positive lag-1 autocorrelation. These properties show that semantic boundaries should be detected by variance-normalized deviations, not by document means or global similarity thresholds. Accounting for autocorrelation leads to a normalized discrete Laplacian detector that identifies significant semantic valleys through a single interpretable decision rule.

Evaluated on five BEIR benchmarks, Seismos consistently improves over fixed-length, recursive, and production semantic chunking baselines under one fixed operating configuration. Without corpus-specific retuning, the same detector transfers to a held-out TREC-COVID corpus. Beyond short documents, Seismos remains competitive with strong fixed-length chunking on long-document (DAPR, ConTEB) and multilingual (MIRACL) retrieval across four encoders, and in end-to-end RAG on LongBench it retrieves better-ranked evidence and yields more faithful answers at 15–23× lower preprocessing cost than production semantic chunkers. Our results suggest that the boundaries needed for effective retrieval are already encoded in embedding-similarity signals, and that principled, efficient, LLM-free chunking can be obtained by detecting them statistically.

Method

Chunking as hypothesis testing

Under the null of no topic shift, the document mean cancels out. The right normalizer is the signal's own variability.

SEISMOS pipeline: encode sentences, compute consecutive cosine similarity, apply the discrete Laplacian, normalize to t_i, split where t_i exceeds k, and index the resulting chunks. — **Pipeline.** Similarity signal → discrete Laplacian → normalized statistic \(t_i\) → split where \(t_i>k\).

Step 1 · Null

The mean cancels

With no boundary at \(i\), the slope into a valley has zero mean, so \(\bar{x}_D\)-scaled thresholds are misspecified.

\[ \operatorname{Var}(x_{i-1}-x_i)=2\sigma^2 \]

Step 2 · Autocorrelation

Similarities are AR(1)

Lag-1 ACF is 0.27–0.38 on every corpus. The second difference has an exact AR(1) variance.

\[ \operatorname{Var}(D_i)=\sigma^2(6-8\rho+2\rho^2) \]

Step 3 · Detector

One self-normalized test

\(\hat{s}_D=\operatorname{std}(D)\) absorbs both \(\sigma\) and \(\rho\). Split where

\[ t_i=\frac{x_{i+1}-2x_i+x_{i-1}}{\hat{s}_D}>k \]

The whole detector in NumPy sketch of Algorithm 2

def seismos_chunk(sentences, encode, k=2.5):
    n = len(sentences)
    if n < 5:
        return [" ".join(sentences)]
    E = encode(sentences)                         # (n, d), L2-normalized
    x = np.sum(E[:-1] * E[1:], axis=1)           # x_i = <e_i, e_i+1>
    D = x[2:] - 2 * x[1:-1] + x[:-2]              # discrete Laplacian
    s = D.std()                                   # absorbs sigma and rho
    B = set() if s == 0 else set(np.flatnonzero(D / s > k) + 1)

    chunks, cur = [], [sentences[0]]
    for i in range(n - 1):                       # boundary i sits between s_i and s_i+1
        if i in B:
            chunks.append(" ".join(cur)); cur = [sentences[i + 1]]
        else:
            cur.append(sentences[i + 1])
    chunks.append(" ".join(cur))
    return chunks

Interactive

Try the detector

A synthetic document with four planted topic shifts (▲) and one gradual trough. Drag k and watch the splits.

k 2.5

null tail Pr(Z>k): 0.6%
splits: 4
shifts found: 4 / 4
noise splits: 0

Signal analysis

Three properties, every corpus

Three SciFact documents: a smooth one with no split, a mixed one where a broad trough is ignored and one sharp valley is split, and a structured one with two sharp valleys giving three chunks. — The same similarity value means different things in different documents, so the threshold must be relative to each document's own signal.

18–23%of documents have sharp topic valleys; 34–58% are flat

> 80%of local minima vanish by prominence τ = 0.10

ρ = 0.27–0.38lag-1 autocorrelation of consecutive similarities

Fraction of local minima surviving as tau increases for five corpora. — Most minima are shallow noise.

Empirical ratio std(D) over std(X) per corpus: 2.03, 1.93, 2.17, 2.09. — \(\hat{s}_D/\sigma_D\) is stable at 1.93–2.17, as the AR(1) model predicts.

Results

One operating point, every benchmark

NDCG@10 (%)	ArguAna	FiQA	NFCorpus	SciFact	TREC-COVID†
LangChain Semantic	63.62	39.18	36.78	73.57	66.90
LlamaIndex Semantic	56.85	35.68	37.05	72.53	66.24
Recursive (best)	59.91	39.35	37.33	74.42	70.00
Fixed (best)	60.62	39.04	36.36	74.22	69.81
SEISMOS · z-score	63.53	39.23	37.63	74.52	71.40
SEISMOS · Laplacian	63.63	39.50	37.63	74.55	71.35

Same encoder (BGE-base) and index for every chunker. † held out from all tuning. Up to +4.98% over the best fixed-length and +6.21% over the best recursive chunker.

	DAPR				ConTEB			MIRACL
nDCG@10	MS MARCO	NQ	CondQA	Genomics	NarrQA	ESG	Geo	JA	DE	AR
Fixed@480	67.1	48.1	27.9	43.2	82.2	46.1	93.6	75.2	48.4	74.3
SEISMOS	68.1	47.1	26.9	42.9	82.2	44.6	93.7	75.6	49.3	74.4
Δ	+1.0	−1.0	−1.0	−0.3	0.0	−1.5	+0.1	+0.4	+0.8	+0.1

Competitive with a strong fixed-length baseline on documents up to tens of thousands of tokens, and across Japanese, German and Arabic with no language-specific tuning.

End-to-end RAG

Better evidence, faster

LongBench NarrativeQA, documents up to 67k tokens. Only the chunker changes.

Chunk nDCG@5

Rank of the evidence-bearing chunk

SEISMOS0.190

Fixed@4800.113

Recursive0.113

LlamaIndex0.108

LangChain0.106

00.050.100.150.20

Preprocessing time

Seconds per ≈27k-token document

LlamaIndex3.105

LangChain2.036

Recursive0.436

SEISMOS0.133

Fixed@4800.0004

00.81.62.43.2

0.687HHEM faithfulness, best of five chunkers (+0.043 to +0.098)

0.952human-rated correctness vs 0.333–0.429 (21 items, 3 annotators)

11 / 100documents with a chunk under 20 tokens, vs 94 / 100 for semantic baselines

What the generator sees

Where does the slime attack Dana and Oscar?

Reference

museum

SEISMOS

The slime attack on Dana and Oscar occurs in the Museum Restoration Studio.

Fixed@480

The slime attack Dana and Oscar in the bathtub at Dana's apartment.

Recursive

The slime attack occurs in the bathtub at Dana's apartment.

LangChain

The slime attack on Dana and Oscar occurs in Dana's apartment bathroom.

LlamaIndex

The slime attack occurs in Dana's apartment when Dana is preparing to bathe Oscar in the old claw-footed bathtub.

BibTeX

@inproceedings{saad2026seismos,
  title     = {{SEISMOS}: A Statistical Signal Detection Framework
               for Semantic Chunking},
  author    = {Saad, Ishtiak Mahmud and Islam, Mominul and
               Ridoy, Shahriyar Zaman and Ahsan, Md Manjurul and
               Wasi, Azmine Toushik},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Abstract

Chunking as hypothesis testing

The mean cancels

Similarities are AR(1)

One self-normalized test

Try the detector

Three properties, every corpus

One operating point, every benchmark

Each correction adds up

Better evidence, faster

Chunk nDCG@5

Preprocessing time

What the generator sees

Reference

SEISMOS

Fixed@480

Recursive

LangChain

LlamaIndex

Reference

SEISMOS

Fixed@480

Recursive

LangChain

LlamaIndex

Reference

SEISMOS

Fixed@480

Recursive

LangChain

LlamaIndex

BibTeX