Semantic Register Compression in Multi-Agent LLM Cascades

Manuele Tele Junior Fernandez

REVIEW 4 major objections 6 minor 12 references

In a three-agent LLM pipeline, the evaluator's rewrite shrinks the distances between label categories in embedding space by up to 28%, while identity passthrough leaves them intact.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An intermediate critic agent in a three-agent LLM pipeline reduces embedding-space separability between label categories by 9–28% across three domains, while identity passthrough preserves it.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A plausible, well-controlled diagnostic for a real risk in LLM cascades, but the headline numbers rest on 10 examples per label and the metric is never tied to decision error. the 4 major comments →

arxiv 2607.14119 v2 pith:U4RBYZ7T submitted 2026-05-12 cs.CL

Semantic Register Compression in Multi-Agent LLM Cascades

Manuele Tele Junior Fernandez This is my paper

classification cs.CL

keywords semantic register compressionmulti-agent LLMinter-label separationembedding geometryAI safety evaluationprompt gradientcascading failures

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to establish a new, measurable failure mode for multi-agent LLM systems: when an intermediate 'Evaluator' agent rewrites a neutral report into critical or evaluative prose, the semantic space of the texts flattens, so labels that were separable at the input become harder to tell apart downstream. After correcting a prompt-language mismatch from an earlier version, the reported compression at the Evaluator stage is 10.3% for fact-checking (LIAR), 28.2% for sentiment (SST-5), and 9.1% for medical triage (Triagegeist), while an identity passthrough preserves separation and a credibility-seeking variant expands it. The paper's controlled variants show that the effect is tied to the specific transformation, not to adding a cascade stage. This matters because a fluent-looking evaluative rewrite can silently remove the very distinctions a downstream Decider needs, and because the paper offers a stage-wise diagnostic—inter-label separation—and a prompt-design mitigation. The central caveat is that all measurements live in one embedding model's geometry and are never validated against actual decision outcomes.

Core claim

Semantic register compression is the claim that an evaluative intermediate rewrite reduces the geometric separability of label groups in a multi-agent LLM cascade. In the corrected v2 runs on a Collector-Evaluator-Decider pipeline, inter-label separation (mean Euclidean distance between label centroids) is flat after Collection and then drops at Evaluation: 0.1480 to 0.1328 for LIAR, 0.2077 to 0.1491 for SST-5, and 0.4433 to 0.4029 for Triagegeist. Five controlled variants localize the cause: identity passthrough preserves separation, critical/balanced/numerical transformations compress it, and a credibility-seeking variant expands it to 0.2052 while biasing labels toward mostly-true/true, s

What carries the argument

Inter-label separation: at each pipeline stage, embed intermediate texts, average the embeddings per ground-truth label to form centroids, and take the mean pairwise Euclidean distance between centroids; compression is the percentage drop relative to the input stage. This metric is the instrument that localizes compression to the Evaluator, separates compression from expansion across evaluator variants, and generalizes the measurement across domains. Its interpretability as information loss rests on embedding-space geometry being a faithful proxy for decision-relevant semantic distinctions.

Load-bearing premise

The load-bearing premise is that distances between label-averaged points in one embedding model measure the semantic distinctions a final Decider actually needs; if that proxy fails, the compression percentages are facts about one embedding model, not evidence of decision-relevant information loss.

What would settle it

Run the same pipeline and train a simple classifier on Collector outputs and on Evaluator outputs: if label recovery accuracy does not drop across the Evaluator, then the geometric compression is not decision-relevant. As a second check, recompute inter-label separation with a different embedding model and with mutual-information estimates; if the reported percentages change qualitatively, they are artifacts of all-MiniLM-L6-v2's geometry.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

A multi-agent pipeline can lose category separability even when every intermediate text reads fluently; compression is invisible at the trace level and must be measured stage-wise.
Safety evaluation should include inter-label separation as a representation-level check alongside final accuracy and trace inspection.
Evaluator prompts can be designed to preserve distinctions: constraint-rich prompts that force evidence, uncertainty, and label-discriminating criteria compress less than a generic 'critique this claim' instruction.
Compression generalizes across domains with different strengths, and domains with larger input separation (such as triage) are proportionally more resistant.
Geometric compression and directional output bias are dissociable: a credibility-seeking evaluator expands separation while still pushing outputs toward mostly-true/true.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The natural safety fix is architectural: let the Decider see the original input or a fused representation, not only the Evaluator's rewrite; otherwise the information lost at the Evaluator is unrecoverable.
The metric should be validated before operational use: a classifier trained on Evaluator embeddings should show a drop in label recovery relative to Collector embeddings, and the percentages should be stable across multiple embedding models and information-theoretic estimators.
The non-monotonic gradient suggests an unexpected prompt-safety implication: balanced, 'neutral-looking' evaluative prompts may be more dangerous than overtly aggressive ones, so prompt guardrails should target register standardization, not just criticality.
A testable extension is to estimate intrinsic dimension of label-conditioned embeddings at each stage; if compression correlates with dimensionality collapse, the mechanism is structural rather than prompt-specific.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

A plausible, well-controlled diagnostic for a real risk in LLM cascades, but the headline numbers rest on 10 examples per label and the metric is never tied to decision error.

read the letter

Short version: the paper gives a crisp, testable description of a plausible failure mode — an intermediate evaluator can make different categories look more alike in embedding space, and the effect depends on the transformation, not just the extra stage. That is worth taking seriously. But the headline percentages rest on 10 examples per label with no error bars, and the metric is never checked against actual decider errors, so the "safety-relevant failure mode" framing outruns the evidence.

What is genuinely good: the controlled design. Identity passthrough is the right control, and the credibility-seeking variant is a smart contrast — it shows you can flip the sign of geometric change with orientation while also shifting outputs. The cross-domain replication across fact-checking, sentiment, and triage gives the phenomenon at least some generality. The paper is also transparent about its own limitations, including a v2 note that corrects a prompt-language mismatch. The authors explicitly call the embedding metric a proxy. No fitted parameters, no post-hoc predictions dressed up as discovery. That earns serious engagement.

The soft spots, in proportion. The biggest is the validity of the metric. Inter-label centroid distance in all-MiniLM-L6-v2 space is never connected to downstream decision accuracy or information-theoretic recoverability; classifier-based validation is deferred to future work. Until that connection is made, the compression numbers are facts about MiniLM geometry, not about information loss the Decider actually suffers. The paper says as much in its limitations, but then continues to call it a failure mode. That is a framing gap, not a hidden flaw.

Second, the main separability numbers come from n=10 per label. No bootstrap, no significance test. Ten points per centroid is very noisy, and the differences (0.1480 → 0.1328) could easily be within sampling noise. The 500-example replication covers only the output distribution, not the separability metric.

Third, there is an unexplained numerical inconsistency: the gradient experiment reports passthrough separation around 0.43, while Table 2 reports the input separation on the same type of data at 0.1480. The revision note doesn't address this. It may be a different sample or a different embedding step, but as written it undermines confidence in the numbers' stability.

The central idea holds up as a hypothesis. This is for people building multi-agent LLM pipelines in safety-sensitive settings and for evaluation researchers. It deserves peer review — the controlled variants and cross-domain replication justify referee time — but it needs a substantially larger sample and an external validation of the metric against decision outcomes before the safety claims can be accepted. I would not cite it in its current form, but I would read the revision.

Referee Report

4 major / 6 minor

Summary. The paper identifies and measures a phenomenon it calls semantic register compression (SRC) in multi-agent LLM cascades. Using a Collector–Evaluator–Decider pipeline, it computes inter-label separation as the mean Euclidean distance between label-conditioned centroids in all-MiniLM-L6-v2 embedding space. It reports that the Evaluator stage compresses this separation in three domains (LIAR: 0.1480→0.1328, 10.3%; SST-5: 0.2077→0.1491, 28.2%; Triagegeist: 0.4433→0.4029, 9.1%), while identity passthrough preserves it and a credibility-seeking variant expands it. A 20-level prompt gradient is reported as non-monotonic, with balanced evaluative prompts producing the largest compression. The authors frame SRC as a measurable, generalizable, and partially preventable failure mode with implications for safety evaluation of multi-agent systems.

Significance. The controlled variants (identity passthrough, evaluator removal, credibility-seeking, balanced, numerical) provide a useful causal skeleton: they suggest that geometric change depends on the specific transformation rather than on the mere presence of an additional stage. The cross-domain replication and the honest revision note documenting a prompt-language mismatch are also strengths. If the geometric metric were validated as a proxy for decision-relevant information, SRC could become a practical stage-wise diagnostic for evaluating information preservation in LLM cascades. However, the paper's central claim currently rests on an unvalidated embedding-space proxy measured on very small samples, so the significance is conditional on the additional validation the authors themselves defer to future work.

major comments (4)

[§2.3, Eq. (2)–(3); §8 Limitations] The central claim that SRC is a 'failure mode' with safety implications depends on the assumption that Euclidean distance between label-conditioned centroids in all-MiniLM-L6-v2 space tracks distinctions the downstream Decider needs. The paper explicitly calls the metric 'a proxy for semantic separability, not an absolute measure of meaning' and defers classifier-based validation to future work. Without validating the metric against Decider accuracy, information-theoretic recoverability, or at least a second embedding model, the compression percentages are properties of one embedding model, not evidence of decision-relevant information loss. This is the load-bearing validity concern and should be addressed before the failure-mode claim is accepted.
[§2.2, Tables 2 and 4] All separability measurements use n=10 per label for LIAR (50 total) and 50 examples per condition for the cross-domain and variant experiments, with no confidence intervals, bootstrap estimates, or significance tests. The 500-example replication in §3 covers only final-label distribution, not inter-label separation. Consequently, the headline quantitative comparisons (10.3% vs 28.2% vs 9.1%; variant differences 0.1180/0.1015/0.1110/0.1418/0.2052) may be sampling noise. The paper needs error bars, a significance test, or a separability replication at larger n before these magnitudes can be interpreted.
[§6 vs §4, Table 2; §5, Table 3] The gradient experiment reports a passthrough baseline with mean separation ≈0.43 and maximum-compression separation ≈0.31, while the main experiments report input separation 0.1480 (Table 2) and 0.1415 (Table 3). This order-of-magnitude discrepancy is not explained in the text. If the gradient used a different sample, a different embedding, or a different computation of separation, then its non-monotonic profile cannot be directly compared with the headline compression numbers. The authors must state the gradient's input separation, sample, and protocol, and reconcile the discrepancy, since the non-monotonicity claim is one of the paper's four contributions.
[§5, Table 3 and revision note] The variant experiment appears to use a different LIAR sample from Table 2 (input 0.1415 vs 0.1480), but this is stated only in the revision note, not in the main text. In addition, the paper does not specify which LLM generated the variant outputs, saying only that the analysis is limited to 'a single primary model' in Limitations. Since the causal claim that transformation type, not cascade presence, governs geometric change rests entirely on this experiment, the authors should state the model, sample, and seeds, and ideally replicate the variant comparison on at least one additional model.

minor comments (6)

[§2.3, Eq. (1)] The centroid equation is rendered awkwardly, with the denominator appearing in a compressed form. Please expand the fraction for readability.
[§3] The single-agent baseline is mentioned as producing a more distributed output, but no table or numbers are provided. Since it is used to argue that collapse is architecture-related, a quantitative comparison would strengthen the claim.
[§4, Table 2 and §5, Table 3] Table 3 reports 'Sep.' but does not give the compression percentage corresponding to each variant, nor confidence ranges. Adding the percentage and a note on the input baseline would make the table self-contained.
[§2.2] The Triagegeist dataset is described as synthetic but no citation or link is provided. Please add a reference or a fuller description so readers can assess its provenance.
[§6] The gradient section states that levels 0–5 remain 'close to the passthrough baseline (mean sep ≈ 0.43)' but does not give the input separation for those runs. Given the order-of-magnitude difference from Table 2, a brief sentence explaining the protocol would prevent confusion.
[§8] The discussions of entropy collapse and intrinsic dimension cite several 2026 preprints. The connection between those theoretical results and the observed non-monotonic gradient is asserted rather than derived; consider making the logical link more explicit or moving these speculations to future-work paragraphs.

Circularity Check

0 steps flagged

No circularity: compression is measured directly from defined inter-label separation, not derived from a fit or from self-citation; the acknowledged proxy limitation is a validity concern, not a circular step.

full rationale

The paper's central derivation is self-contained and non-circular under the requested criteria. Semantic register compression is defined directly as the percentage reduction in measured inter-label separation, C_t = (S_input − S_t)/S_input × 100, where S_t is computed from label-conditioned centroid distances in a fixed embedding space. There are no fitted parameters, no hidden regressions, and no 'prediction' that reduces to a fitted input. The five Evaluator variants are empirical comparisons against the input baseline, and the identity-passthrough and Evaluator-removed controls test whether compression is attributable to the transformation rather than to mere cascade depth. The 20-level gradient is an independent exploration, and the cross-domain comparison uses identical prompts and reports measured separation values. No load-bearing step depends on a citation, and no cited uniqueness theorem or prior work by the authors is invoked to force the conclusion. The paper explicitly labels its embedding metric 'a proxy for semantic separability, not an absolute measure of meaning' (Limitations, §8), so whatever concern exists is about construct validity — whether embedding geometry tracks decision-relevant information — not about circular derivation. The skeptical observation that the metric used to define SRC is also the metric used to confirm it describes the operational definition of the phenomenon, not a tautological inference from an input to an output. The numerical inconsistency between the gradient's 'mean sep ≈ 0.43' and Table 2's input separation of 0.1480 is a reporting or scale discrepancy, not evidence of circularity. Therefore no circular step is identified, and the honest finding is a score of 0.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 1 invented entities

The central claim rests on domain assumptions about embedding-space geometry tracking decision-relevant information, sample stability, and model/dataset representativeness. There are no fitted free parameters; the compression percentages are direct measurements. The main construct, semantic register compression, is operationally defined by the paper's own metric, so external validation of the metric against downstream decisions is the key missing check.

axioms (4)

domain assumption Inter-label centroid distance in all-MiniLM-L6-v2 embedding space is a valid proxy for decision-relevant semantic separability.
Invoked in §2.3 and §8; never validated against downstream accuracy or information-theoretic recoverability; the paper itself calls the metric 'a proxy for semantic separability, not an absolute measure of meaning.'
domain assumption The Collector–Evaluator–Decider architecture and GPT-4.1-family models are representative enough of multi-agent LLM systems to support a generalizable failure-mode claim.
§2.1, §3; only OpenAI gpt-4.1 variants are tested, and the evaluator-variant analysis uses a single primary model.
domain assumption Label-conditioned centroid means from n=10 examples per label are stable enough to report compression percentages to one decimal place.
§2.2, §4; no variance or CI is reported; the 500-example replication only tests the output-label distribution, not the separability metric.
domain assumption Triagegeist synthetic emergency-department data is representative of real clinical triage text.
§2.2; acknowledged as synthetic in Limitations, so the triage generalization claim assumes this representativeness.

invented entities (1)

Semantic register compression (SRC) independent evidence
purpose: Named failure mode: reduction of inter-label separability in embedding space caused by intermediate evaluative rewriting in multi-agent cascades.
The paper provides a public measurement recipe (inter-label separation) that other labs can apply, so the construct is falsifiable; however, its safety relevance is not independently validated beyond the embedding proxy.

reviewed 2026-08-02 · how reviews work

0 comments

Cite this review

Pith. "Pith review of Semantic Register Compression in Multi-Agent LLM Cascades." pith.science (2026). https://pith.science/paper/U4RBYZ7T

@misc{pith2026260714119,
  author       = {Pith},
  title        = {Pith review of: Semantic Register Compression in Multi-Agent LLM Cascades},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U4RBYZ7T}},
  note         = {Machine review of arXiv:2607.14119}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

Multi-agent LLM systems commonly decompose complex tasks into specialized roles. However, this modularity introduces a representational risk: when intermediate agents transform text across linguistic registers, they can systematically compress the semantic distinctions needed for accurate downstream decisions. We term this phenomenon semantic register compression and characterize it as an observable failure mode in multi-agent cascades. Using a three-agent pipeline (Collector-Evaluator-Decider), we quantify compression via inter-label separation in sentence-transformer embedding space. Across political fact-checking (LIAR), sentiment analysis (SST-5), and medical triage (Triagegeist), critical evaluation reduces label separability at the Evaluator stage, while identity passthrough preserves it nearly fully. Five controlled variants show that geometric change depends on the specific intermediate transformation rather than on the mere presence of an additional cascade stage. A credibility-seeking variant expands rather than compresses inter-label separation, while shifting outputs toward mostly-true, demonstrating that transformation valence controls both the direction and the sign of geometric change independently of compression magnitude. Compression generalizes across the three domains with domain-dependent intensity (10.3% in fact-checking, 28.2% in sentiment, 9.1% in triage). A 20-level prompt gradient reveals a non-monotonic compression profile: balanced evaluative prompts produce the strongest compression, while extreme critical prompts show irregular moderate compression. These results demonstrate that semantic register compression is a measurable and generalizable phenomenon in multi-agent LLM systems, with implications for safety evaluation in high-stakes domains.

Figures

Figures reproduced from arXiv: 2607.14119 by Manuele Tele Junior Fernandez.

**Figure 1.** Figure 1: Semantic register compression manifests as distributional collapse. (A) The LIAR input sample is balanced across five sampled veracity categories (20% each, n = 10 per label). (B) The baseline multi-agent pipeline collapses toward central categories: half-true (48%) and barely-true (42%) dominate, while mostly-true and true are never selected (0%). The labels mostly-true and true are never selected in the… view at source ↗

**Figure 2.** Figure 2: Inter-label separation quantifies geometric compression across pipeline stages. The baseline architecture shows a sharp reduction in interlabel separation at the Evaluator stage, while the identity passthrough control maintains stable separation across stages. Numerical values are reported in [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 1 canonical work pages

[1]

2017 , publisher =

Wang, William Yang , booktitle =. 2017 , publisher =. doi:10.18653/v1/P17-2067 , url =

doi:10.18653/v1/p17-2067 2017
[2]

2019 , publisher =

Reimers, Nils and Gurevych, Iryna , booktitle =. 2019 , publisher =. doi:10.18653/v1/D19-1410 , url =

doi:10.18653/v1/d19-1410 2019
[3]

and Ng, Andrew Y

Socher, Richard and Perelygin, Alex and Wu, Jean and Chuang, Jason and Manning, Christopher D. and Ng, Andrew Y. and Potts, Christopher , booktitle =. 2013 , publisher =. doi:10.18653/v1/D13-1170 , url =

doi:10.18653/v1/d13-1170 2013
[4]

and Mordatch, Igor , year =

Du, Yilun and Li, Shuang and Torralba, Antonio and Tenenbaum, Joshua B. and Mordatch, Igor , year =. doi:10.48550/arXiv.2305.14325 , url =. 2305.14325 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2305.14325
[5]

doi:10.48550/arXiv.2503.13657 , url =

Cemri, Mert and others , year =. doi:10.48550/arXiv.2503.13657 , url =. 2503.13657 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2503.13657
[6]

and Eskridge, Thomas and Feltovich, Paul J

Bunch, Larry and Bradshaw, Jeffrey M. and Eskridge, Thomas and Feltovich, Paul J. and Lott, John and Uszok, Andrzej and Carvalho, Marco , booktitle =. 2012 , publisher =. doi:10.1109/WI-IAT.2012.272 , url =

work page doi:10.1109/wi-iat.2012.272 2012
[7]

, year =

Mabrok, Mohamed A. , year =. doi:10.48550/arXiv.2603.22301 , url =. 2603.22301 , archivePrefix =

doi:10.48550/arxiv.2603.22301
[8]

doi:10.48550/arXiv.2601.03779 , url =

Baroni, Marco and Cheng, Emily and de-Dios-Flores, Iria and Franzon, Francesca , year =. doi:10.48550/arXiv.2601.03779 , url =. 2601.03779 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2601.03779
[9]

doi:10.48550/arXiv.2603.26993 , url =

Ao, Ruicheng and Gao, Siyang and Simchi-Levi, David , year =. doi:10.48550/arXiv.2603.26993 , url =. 2603.26993 , archivePrefix =

doi:10.48550/arxiv.2603.26993
[10]

doi:10.48550/arXiv.2603.04474 , url =

Xie, Yizhe and Zhu, Congcong and Zhang, Xinyue and Zhu, Tianqing and Ye, Dayong and Qi, Minfeng and Chen, Huajie and Zhou, Wanlei , year =. doi:10.48550/arXiv.2603.04474 , url =. 2603.04474 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2603.04474
[11]

doi:10.48550/arXiv.2604.13349 , url =

Li, Yiping and An, Zhiyu and Du, Wan , year =. doi:10.48550/arXiv.2604.13349 , url =. 2604.13349 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2604.13349
[12]

doi:10.48550/arXiv.2602.18450 , url =

Alpay, Faruk and Kilictas, Bugra , year =. doi:10.48550/arXiv.2602.18450 , url =. 2602.18450 , archivePrefix =

doi:10.48550/arxiv.2602.18450

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.

[1] [1]

2017 , publisher =

Wang, William Yang , booktitle =. 2017 , publisher =. doi:10.18653/v1/P17-2067 , url =

doi:10.18653/v1/p17-2067 2017

[2] [2]

2019 , publisher =

Reimers, Nils and Gurevych, Iryna , booktitle =. 2019 , publisher =. doi:10.18653/v1/D19-1410 , url =

doi:10.18653/v1/d19-1410 2019

[3] [3]

and Ng, Andrew Y

Socher, Richard and Perelygin, Alex and Wu, Jean and Chuang, Jason and Manning, Christopher D. and Ng, Andrew Y. and Potts, Christopher , booktitle =. 2013 , publisher =. doi:10.18653/v1/D13-1170 , url =

doi:10.18653/v1/d13-1170 2013

[4] [4]

and Mordatch, Igor , year =

Du, Yilun and Li, Shuang and Torralba, Antonio and Tenenbaum, Joshua B. and Mordatch, Igor , year =. doi:10.48550/arXiv.2305.14325 , url =. 2305.14325 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2305.14325

[5] [5]

doi:10.48550/arXiv.2503.13657 , url =

Cemri, Mert and others , year =. doi:10.48550/arXiv.2503.13657 , url =. 2503.13657 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2503.13657

[6] [6]

and Eskridge, Thomas and Feltovich, Paul J

Bunch, Larry and Bradshaw, Jeffrey M. and Eskridge, Thomas and Feltovich, Paul J. and Lott, John and Uszok, Andrzej and Carvalho, Marco , booktitle =. 2012 , publisher =. doi:10.1109/WI-IAT.2012.272 , url =

work page doi:10.1109/wi-iat.2012.272 2012

[7] [7]

, year =

Mabrok, Mohamed A. , year =. doi:10.48550/arXiv.2603.22301 , url =. 2603.22301 , archivePrefix =

doi:10.48550/arxiv.2603.22301

[8] [8]

doi:10.48550/arXiv.2601.03779 , url =

Baroni, Marco and Cheng, Emily and de-Dios-Flores, Iria and Franzon, Francesca , year =. doi:10.48550/arXiv.2601.03779 , url =. 2601.03779 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2601.03779

[9] [9]

doi:10.48550/arXiv.2603.26993 , url =

Ao, Ruicheng and Gao, Siyang and Simchi-Levi, David , year =. doi:10.48550/arXiv.2603.26993 , url =. 2603.26993 , archivePrefix =

doi:10.48550/arxiv.2603.26993

[10] [10]

doi:10.48550/arXiv.2603.04474 , url =

Xie, Yizhe and Zhu, Congcong and Zhang, Xinyue and Zhu, Tianqing and Ye, Dayong and Qi, Minfeng and Chen, Huajie and Zhou, Wanlei , year =. doi:10.48550/arXiv.2603.04474 , url =. 2603.04474 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2603.04474

[11] [11]

doi:10.48550/arXiv.2604.13349 , url =

Li, Yiping and An, Zhiyu and Du, Wan , year =. doi:10.48550/arXiv.2604.13349 , url =. 2604.13349 , archivePrefix =

Pith/arXiv arXiv doi:10.48550/arxiv.2604.13349

[12] [12]

doi:10.48550/arXiv.2602.18450 , url =

Alpay, Faruk and Kilictas, Bugra , year =. doi:10.48550/arXiv.2602.18450 , url =. 2602.18450 , archivePrefix =

doi:10.48550/arxiv.2602.18450

Semantic Register Compression in Multi-Agent LLM Cascades

The reading

Core claim §

What carries the argument

Load-bearing premise §

What would settle it §

If this is right

Where Pith is reading between the lines

discussion (0)

Core claim

Load-bearing premise

What would settle it