acceptodds
Under review as a conference paper at ICLR 2027

One Sentence, Two Meanings: Semantic Multiplexing for LLM-Generated Text Steganography

Abstract

Semantic steganography typically hides information by generating LLM text via mapping secret bits to a single semantic space. A natural text, however, can readily convey two distinct meanings rooted in different domains of knowledge through a single carrier. In this work, we attempt to transform this polysemy of natural language into a measurable semantic steganographic channel resource, called semantic multiplexing. We introduce Iphis, a steganographic protocol that embeds two independent secret bit streams in one carrier. For knowledge bases such as urban traffic and computer networks, there are homographs like "traffic" and "path", called pivot words. Iphis introduces joint semantic coordinates to encode the semantics from two different spaces, and uses pivot words to assist LLM in generating carrier sentences. Text generated by Iphis has two layers of semantics: the surface semantics can be naturally understood, while the inner ones can only be extracted when corresponding knowledge base is revealed. Our evaluation indicates that the second semantic stream doubles the matched payload per carrier, initially demonstrating potential of semantic multiplexing. The evaluation also shows that a warden without key cannot recover hidden meaning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Generative Steganography, Large Language Models, Semantic Multiplexing, Controllable Text Generation, Rejection Sampling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9e3mOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
1(−1)
Presentation
1
Soundness
1(−1)
Read review
Paper summary

This paper proposes Iphis, a text steganography protocol that uses pivot words to encode two semantic readings in one text carrier. Experiments on a synthetic semantic channel and a real-LLM pipeline evaluate payload capacity, stream independence, naturalness, and detectability under a specified detector suite.

Critical strengths
  1. Novel problem formulation. The paper treats polysemy as a resource for encoding two semantic streams in one carrier, extending the conventional single-stream setting.
  2. Evaluation of key protocol properties. The authors test the method on both a synthetic channel and a real-LLM pipeline. Per-stream key-scrambling experiments provide evidence that decoding one stream can fail while the other remains recoverable.
Critical weaknesses
  1. Scalability beyond two streams remains unclear. The main experiments focus on the dual-stream setting. In the higher-width boundary experiments, retry-exhaustion rates reach 85.9%–94.8% at q=4q=4, raising concerns about the sampling cost of extending the protocol.

  2. Dependence on a curated semantic space. The protocol uses a synthetic knowledge base with hand-annotated pivots and is evaluated with one model in one language. Its applicability to broader, naturally occurring semantic resources remains unclear.

  3. Limited robustness to transformations. The paper states that repetition does not address paraphrasing. It also does not evaluate a warden adaptively trained against the protocol, so the security results should be interpreted within the tested settings.

Questions for authors
  1. Can the protocol support three or more streams? Would this require a separate semantic domain for each stream, or could multiple streams share a semantic space? What trade-offs in decoding reliability and rejection-sampling cost would this introduce?

  2. What role does LLM understanding of puns play? Since Iphis uses a predefined semantic space and extraction rules, how does its protocol-defined ambiguity differ from an LLM’s contextual interpretation of natural puns?

  3. How should “independent keys” be understood? The paper reports per-stream key-scrambling experiments, while the protocol pseudocode refers to a shared key KK. Could the authors clarify whether the streams use separate keys, derived keys, or independent key interventions only for evaluation?

  4. Could a multi-stream design still use a surface/inner-layer structure? For example, could one reading remain the surface interpretation while additional streams are encoded as separate inner readings? What changes to the extraction rules would be required?

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9e3m - Part 1

Thank you for the review which we have read again and again, as a set of directions rather than only assessment. The concerns mentioned critical weakness and questions are where this line of work has to go next, and they are useful for us to improve the submission. We tried our best to response your opinions and questions in order(including a combination for Q1 and Q4). We also report the experiments we could run during the rebuttal.

Critical Weakness 1: scalability beyond two streams

Thank you for this insightful question. We decompose the discussion of scalability into two levels, a third stream(even more) and a wider label. Both of the issues were discussed during the research stage, but not so thorough. And your review prompted us to revisit it.

A third stream means introducing a third reading into the sentences. That is to say, building a triple semantic graph. Although the dual semantic graph of Iphis was constructed manually, it still adhered to the pivot words inherent in natural language. Triple semantic, by the contrast, is extremely rare in natural language. Thus the obstacle to a third stream would be the resource, rather than the sampler.

On the other hand, we did report the retry exhaustion for a wider label q = 3 and q = 4, which seemed impossible to construct a wider label. We found the bottleneck is accpetance rate: under blind rejection a candidate matches with probability 2^-q, and a frame succeeds only if all N positions succeed, so the cost compounds. The measurement in the submitted manuscript was only the average of single attempt, thus it only reflected the success rate under certain maximum retries(r_max = 96).

Samplerr_maxnonce retriesq = 3q = 4
Blind rejection (submitted)9610.9920.116
Blind rejection60011.0000.999
Enumeration + P0-weighted draw—11.0000.999
Blind rejection9641.0000.452
Blind rejection60041.0001.000
Enumeration + P0-weighted draw—41.0001.000

The tablet above shows measurements during this discussion. Frame success at the finite-alphabet boundary, rebuilt offline with the same channel, label function and codec (256 physical positions, five seeds × 200 frames = 1,000 frames per cell). Rows with nonce retries = 1 are the submitted configuration; rows with 4 are the protocol's nonce retries. The submitted boundary reports 991/1000 and 982/1000 at q = 3 and 141/1000 and 52/1000 at q = 4 under a single attempt, which our rebuild reproduces (0.992 and 0.116). "Enumeration" means the encoder lists the legal coordinates and draws among the label matches with P0 weights, which has the same distribution as rejection sampling but no 2^-q draw cost.

The results show that a wider label are feasible, but how to reduce the cost to a reasonable level is another issue to be considered.

AuthorsOpenReview ↗

Response to Reviewer 9e3m - Part 2

Critical Weakness 2: applicability to broader naturally occurring semantic resources

Thank you for pointing out this critical gap. We do agree that the coverage of the semantic channel decides how far the protocol can go. We considered several possible extensions during the research stage. However, as the initial step of a line of work, we submitted the best we can do at the current time, to implement the protocol itself with available resource at that time.

About model. For the main closed loop we used a single renderer, deepseek-chat with reasoning off. The receiver side is already checked across three models, since Qwen3-30B-A3B(2/2 frames) and DeepSeek-V3.2(1/2) re-decode the B stream from frozen carriers. During the rebuttal we added a second renderer, Qwen3-14B, which closes the loop at 10/10 frames, and a third, GLM-4-9B, which fails to encode at 0/10. The failure is in the render-and-extract round trip, not in the sampler. Due to the lack of a stable and compliant source of services, we did not use GPT and Claude.

About language and scale. In the early stage of the research, we did consider building a bilingual knowledge base and found the English pun benchmark SemEval Task 7. However, after carefully checking, we found that it could not be directly used for the experiment. (Another potential source of multilingualism is the "Kakekotoba" in Japanese rhetoric, but there still lacks of available constructions.) Moreover, the English knowledge base requires additional effort to handle possible language-specific problems. Therefore, this work was temporarily put on hold. Similarly, the larger language database stopped at the diffuculties of construction basis. We found a terminology database (TermOnline) that could potentially serve. But batch access to TermOnline requires an agreement and we were unable to finish the signing process at that time.

We are now going to extent our work from three aspects. More domain pairs, chosen to span different degrees of structural overlap rather than just more of the same. An English knowledge base, to separate what is language-specific from what is not. And a renderer spectrum rather than a single model, since our own runs show that encodability is renderer-dependent. All the extension should be by selection rather than by volume. Each added unit has to be checked for whether the renderer can express it and whether the receiver can extract it, and units that fail that check are not useful even if they are semantically valid.

Critical Weakness 3: robustness to transformations and adaptive warden

Thank you for raising this critical issue for practical steganography. We sincerely accept the critics. Then we will report the attempts we have made and what we found useful to improve the work.

Robustness to rewriting. Our submission contains the ablation for repetition, which is an error-tolerance mechanism, nothing to do against rewriting. The difficulty of resisting rewriting attack is that the cue is lexical and appears once. Rewriting of pivot words will exactly remove it. The potential solution is push the cue toward structure and pragmatics, instead of vocabulary. But now our extractor has not been able to read it yet. The improvement direction includes structure cues, a second cue of different type in the same sentence and a bit-error curve against attack strength.

Adaptive warden. During the discussion, we conducted some additional experiments to determine the separability of carriers. Against news and Wikipedia text(500 sentences each) the carriers separate at 0.952 and 0.888, and two topic-matched covers at about 0.98. For more precise attribution, we add a plain text control, which seperates at 0.808. So most of the 0.952 is the renderer, and the second reading contributes about 0.14. Our attempts to suppress it never got below 0.916 against a target of 0.85. Our interpretation is that the pivot is itself unusual in this genre: a news sentence does not normally contain the vocabulary containing the second reading, so the carrier is weird before it encodes anything. An adaptive warden would learn exactly that weirdness, which is why we think the answer is a renderer that can carry a second reading in words the genre would use anyway, not a better key. However, all is not lost. The example of rhymes proposed in §1 proved that proper genre can absorb anomaly. "Damp green thoughts were screaming on the bed" appearing in an ICLR 2027 submission could be regarded as a failed generation, but a fantastic verse in avant-garde poem.

We are looking forward for the next experiments as follows: an adaptive warden with a renderer-matched control, and a construction whose second reading rides on words the genre would use anyway.

AuthorsOpenReview ↗

Response to Reviewer 9e3m - Part 3

Answer to Question 1 and 4

First of all, please allow us to give the general conclusion: a third or more streams are possible in principle, but the real bottleneck is the semantic graph construction.

For semantic spaces, a stream is a reading, which relies on a hypothesis plus a set of cues the receiver can verify. Streams with disjoint cue sets can share one semantic space. A separate domain per stream is one way to guarantee that, not a requirement. If two streams share one coordinate space and one sentence with proper segmentation, they can be separated by which reading's cues fire.

For trade-offs, each added bit per position grows the label, and under blind rejection the acceptance per position falls as 2^-q while a frame needs all N positions, so at q = 4 only about one frame in eight survives. We measured the retry cost by enumerating the legal coordinates and raising r_max, and finally restored 999/1000 frames. So what binds is the size of the resource (every target bucket must be non-empty, N ≳ 2^q ln(1/ε)) and decoding reliability, since each added reading brings its own extraction error and must be told apart from the others.

However, semantic space segmentation and reject-sampling costs are not the only obstacles for multistream extension. The weirdness of such generated text will result in extra fragility to adaptive warden, as we already answered in the discussion for Critical Weakness 3. It shares a lot related to Q4, so we combine the two responses here.

For Q4. There are two methods to add the third stream. The first one is to "stack" another layer of reading on the second meaning, leading to a three-way polysemy of one pivot. However, the triple semantics are extremely rare in natural language and the possibility of construction remains purely theoretical. The second approach is parallelly embedding two readings into the same inner layer, each with its own independent pivot. It sounds more possible. But the real cost is weirdness. Each inner reading brings its own pivot vocabulary, which makes the sentence looks odd. Carriers with two kinds of weirdness will be even worse. Whether there exists a genre can absorb that much weirdness remains an open question. It is definitely a difficult construction problem.

Answer to Question 2

Natural puns are indeed where our journey began. The very first construct of this study emerged from a discussion on translating natural puns(from Japanese to Chinese). However, finally the protocol does not ask the model to understand a pun. Because a pun is one kind of dual meaning, not the only one: the same ambiguity carries metaphor, irony and the oracle, and humor is only one of the effects it can serve. The construction of Iphis realises dual readings in a single sentence based on the pair (a, b) and the context, leaving a cue the extractor can use. Extraction is rule-based, leaving nothing to do with "understanding".

This is also why encodability depends on the renderer rather than on the protocol. Qwen3-14B closes 10/10 frames, while GLM-4-9B fails at 0/10 with carrier extraction mismatching. What Iphis depends on is cue realisation, not interpretation.

In the paper we have made a contrast: "for them the second reading is an interpretation to be judged; for Iphis it is a state fixed by the candidate space and recovered under the key." The second reading of natural pun is open-ended. It is whatever the hearer makes of it, it is not tied to a declared unit, and there is nothing to authenticate. In Iphis what we want is a knowledge-gated information hiding, where the second reading can be extracted, keyed and authenticated. Dual meaning is a channel rather than a rhetorical effect. The literary fluency, innovation and humor can enhance stealthiness, but are not essential.

Answer to Question 3

Thank you for pointing out this mistake. It is our fault that caused this inconsistency. The pseudocode in submission gives one master key and one PRF, with both label bits taken from the same digest, which is consistent with the code in the anonymous repository. §4.6 and the appendix describe per-side keys. The protocol path never implemented that derivation and the side keys were only used in the audit path. We should have checked it to match the code, and we are not defending the wording.

During this discussion period we have changed the implementation so that each stream gets its own key derived from the master key. With that we conducted some experiments. The result shows that scrambling one side destroys only its payload (BER 0.502 for B, 0.490 for A) while the other side is preserved bit for bit. Isolation is now a property of the construction rather than a path in the audit code.

This change does not move the numbers: acceptance is 0.4974 at q = 1, 0.2458 at q = 2, and 0.2459 for the matched single stream at q = 2, against 0.498 / 0.246 / 0.246 in the paper.

Reviewer_Hz9AOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a steganographic encoding protocol, Iphis, which encodes two independent information streams into a single carrier.

It gives the Iphis algorithm, and demonstrates that encodings can be recovered from the messages.

Critical strengths

None

Critical weaknesses

W1: The paper is poorly written, and shows clear signs of significant LLM usage in writing. Although the authors acknowledge that they "used generative AI tools [to] polish the manuscript", there is reason to suspect that AI was used to draft large portions of the manuscript which were not subsequently edited. As one example, App. A.5 includes "requires the cover corpus asserted by experiments/run_coordinate_warden.py; the repository holds several pivot cover sets and only one of them is the corpus Table 3 consumes." It is highly unusual to give individual scripts within a paper body, but in my experience is the kind of thing frequently done by LLMs when drafting write-ups. Likewise, the Reproducibility Statement boasts "The codebase only depends on the Python standard library and is covered by unit tests." What relevance does the unit test coverage have to the paper's scientific merits? Likely none, but is it the kind of thing an LLM would proudly declare. Moreover, the main text contains many stylistic choices which are indicative of LLM writing, for example a frequent use of colons. Finally, when I insert the first two paragraphs of a randomly sampled section (3.1) into the AI detector Pangram, it flags as 100% AI written. While using LLMs to write is not inherently reason for rejection, the more important issue is that the text is extremely dense, poorly structured, and difficult to follow. Objects are used with no introduction or explanation (e.g., "prompt v2" being used on page 5), and the authors seem to struggle to convey the ideas in a way that could be followed by a reader not already familiar with them. Regardless of AI involvement, such issues would need to be thoroughly addressed for this work to be published.

W2: Prompts are not given in full, but only as "English gloss". This is strange choice, also indicative of LLM involvement. Why not give the actual prompts? Are these the actual prompts? What does "gloss" mean here?

W3: In what sense is the audit A.3 "Pre-registered"? Do the authors have a link to an public archive where they pre-registered the investigation?

W4: The method seems to significantly underperform the baselines. The authors admit this themselves, in 4.6. Further, comparison against existing methods is obfuscated and not made clear within the main text.

W5: The problem setting is undermotivated. The authors seem to see their primary contribution as encoding two payloads within a single information stream. It is not clear why this is a problem one should care about, and the authors do not motivate or explain this. Further, the authors do not seem to make an attempt to adapt existing baselines to this problem statement, which allows them to claim their numbers are "not commensurable".

Questions for authors

Q1: Can you clarify how LLMs were used in drafting the manuscript? Was involvement limited to "polishing", or were LLMs used to draft portions of the text itself? Can you account for why portions of your manuscript flag as 100% AI written?

Q2: Are there ways that one can adapt simple baselines to your problem setting? If so, how do they fare?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Hz9A

Thank you for the reviewing. We response the most serious point first: the manuscript was written by us, paragraph by pragraph. We have declared, in the AI use statement, that we use AI tools for the following purposes: sourcing information, building the synthetic knowledge bases, assisting code implementation, giving feedback on experiments, and polishing the manuscript. We have already stated on text, following the author guideline. The initial idea of Semantic Multiplexing originated from a discussion about the translation of puns(from Japanese to Chinese), not any LLM. The examples of rhymes we presented in §1 were also one of the sources of inspiration. We shared it and hoped it would help readers understand this work. We did ask LLM questions such like "is this expression appropriate in English?" to polishing the manuscript. We accept the readability criticism, text too dense, objects used before explained — but poor writing and ghostwriting are fundamentally different.

There are indeed some flaws in the submission, which we should response one by one. The internal script path in Appendix A.5 and the sentence about unit-test coverage should not have been in the paper. "English gloss" was a poor way to present a Chinese prompt that we should have given verbatim. (This was a slip, as we initially aimed to ensure that the presentation of prompts remained consistent with the language of the paper.) "Pre-registered" was too strong and there is no public archive. And we should have used "pre-specified".

About the baselines. It is unproper that we claimed our numbers were "not commensurable" without adapting anything, leading to an obfuscated comparison. The adapted baseline is a matched single stream at the same symbol width, on the same channel and position budget. During the rebuttal we ran it, and it ties with the dual stream on rate, 1.0 bit per carrier, acceptance 0.246. The second stream does not add rate but the separability, which we should have written in the main body of paper.

I appreciate the authors' engagement with the criticism. I stand by my assertion that large portions of the text were LLM generated (given that they were flagged by AI detection systems which have very low false-positive rates [1]), but I will allow the Area Chairs to make up their own mind on this issue, since it does not seem feasible to settle the matter one way or another.

Regardless, the more substantive point is that the text is extremely hard to parse and difficult to read (a criticism that the authors accept). I appreciate their willingness to change individual faults and issues that I pointed out. However these were just examples of a much broader and systematic issue with the paper's writing, and dealing with these would not fix the root of the issue.

I am maintaining my score.

[1] https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5407424

Reviewer_Mq4wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces semantic multiplexing, a semantic steganography paradigm that aims to encode two independently keyed bit streams in a single natural-language carrier. The proposed protocol, Iphis, combines a public semantic channel, keyed rejection sampling, and an LLM-based pivot renderer to generate sentences with two domain-specific readings. The authors evaluate the protocol on a synthetic cross-domain semantic graph and conduct a limited end-to-end evaluation using a real LLM. They report a matched net payload increase from 0.5 to 1.0 bit per carrier, alongside key-scrambling experiments and steganalysis evaluations. The paper presents an interesting perspective on exploiting linguistic ambiguity for information hiding. Its protocol is described in considerable detail, and the independent decoding of two streams is supported by controlled experiments. However, the current evaluation leaves substantial uncertainty about the practical security, generality, and efficiency of the proposed approach. In particular, the semantic channel is highly constrained, the real-LLM evaluation is small, and the steganalysis results do not establish robust indistinguishability against a capable adversary. I therefore find the central idea promising but the empirical evidence insufficient to support the broader claims.

Critical strengths

The idea of exploiting linguistic polysemy to encode two independently keyed messages in one carrier is conceptually interesting. Rather than simply increasing the number of bits encoded in a single semantic representation, the paper explicitly separates two semantic readings and attempts to make them independently recoverable. This provides a potentially useful perspective on semantic text steganography. Iphis integrates keyed rejection sampling, pivot-based text generation, replay verification, repetition coding, and stream-level authentication. The key-scrambling experiments are particularly useful: scrambling one stream disrupts its recovery while preserving the other stream. This provides evidence that the two streams are not merely duplicated representations of the same payload.

Critical weaknesses
  1. The reported capacity increases from 0.5 to 1.0 bit per carrier, but q also changes from 1 to 2, making it difficult to attribute the gain specifically to dual-stream decoding. Moreover, the proposed benefits, such as hierarchical messaging and rate adaptation, are not experimentally validated.
  2. The cover texts are generated using a matched pipeline rather than drawn from natural text, so the evaluation does not establish whether steganographic texts are distinguishable from ordinary language. The authors also do not evaluate wardens targeting artifacts introduced by semantic-pivot constructions, leaving practical stealthiness insufficiently demonstrated.
  3. The method relies on a manually constructed semantic graph with predefined domains, pivot words, and contexts. It remains unclear whether the approach generalizes to other domains, languages, or less constrained settings. Evaluations on independently constructed semantic channels are needed to substantiate its generality.
Questions for authors
  1. What is the effective payload rate after accounting for framing, authentication, repetition, failed frames, and the full generation cost?
  2. Given the limited raw capacity, what practical use cases specifically benefit from two independently keyed semantic streams rather than a conventional single-stream scheme?
5 comments
AuthorsOpenReview ↗

Response to Reviewer Mq4w - Part 1

Thank you for your detailed summary of this manuscript and valuable comments. We have rechecked the experimental data records used in the submission, conducted some additional experiments as much as we can, and updated our views by integrating some progress of the ongoing follow-up work. We now organize these contents as follows.

Critical Weakness 1: attribution of capacity gain

First of all, please allow us to explain why the comparison appears as it does in the submitted version. We did not overlook it, but could not run that configuration at that time.

The ablation experiment we wanted was one stream at the same symbol width. But it could not run. The encoder accepted a two-bit single stream, but the decoder read one bit per side per position, so the frame failed at authentication. We changed the design and it worked, which was dual stream at q = 2 against a single stream at q = 1. We acknowledge that it was somehow a kind of compromise which changed two factors at once.

Following your advice, we revisited the decoder. The decoder now votes on each bit plane separately, and the missing cell runs, a single stream at q = 2. The result shows that R_net is 1.0, acceptance 0.246, and 4.07 attempts per position, identical to the dual stream. So the change from 0.5 to 1.0 is due to q alone.

We also want to explain why we consider the success of dual semantic streams as a meaningful result, although their rate is not so high. The seperately authenticated and keyed second payload mainly refer to a kind of conceptual isolation. The two streams can be decoded by different receivers. If one stream is destroyed, the other can still arrive. Scrambling one side leaves the other side intact (BER 0.502 vs 0.490). This is why we believe that semantic multiplexing will ultimately gain advantages and it is also the initial motivation for our research.

Critical Weakness 1+: hierarchical messaging / rate adaptation

For the sake of caution, we must state that neither "hierarchical messaging" nor "rate adaptation" appears in this submission. So we understand them as your summary after reading the manuscript.

If what you mean by hierarchical messaging is the two-level disclosure we were aiming at, §4.4 and Appendix A.1 have shown the isolation between the two levels. A receiver holding K_B and the B-side vocabulary recovers the second payload but cannot authenticate the first without K_A. The surface meaning is understandable to any reader (but it does not mean that everyone can extract the secret information). That is the hierarchy we built: a surface reading protects a knowledge-gated inner one.

On the other hand, we did not make clame on rate adaption. If you feel that the paper suggested it, it should be due to the deficiency of our writing. We actually reported the opposite evidence: acceptance falls as 2^-q, retry exhaustion is 0.9–1.8% at q = 3 and 85.9–94.8% at q = 4, and the payload is set by the frame length we choose (1 B → 240, during the rebuttal period we also conducted additional 8 B → 408) rather than by adapting a rate to a channel. Thank you again for the valuable oppinion.

AuthorsOpenReview ↗

Response to Reviewer Mq4w - Part 2

Critical Weakness 2: cover generated from matched pipeline

Please allow us to explain the design goal of the evaluation at first. At the very beginning, we merely wanted to determine whether the key leaves a trace. To make sure that question, we fixed the renderer with the same model, same frozen prompts, same replay gate, while removed the key label. That is, as you have pointed out, the matched pipeline cover. The result showed that key and sampler did not add extra signature, for the nine detectors sit at 0.50 with the label removed. But it cannot be used to determine whether carriers are similar to original text. We should have specified the scope of statement more clearly in the manuscript.

Thank you for pointing out the missing important half. During the rebuttal stage, we have conducted an additional experiment. We compare the 2400 published carriers in the submission against natural ones from two sources: 500 sentences of Chinese government and transport news and 500 Chinese Wikipedia sentences, length-stratified to the carrier distribution. The detectors are a cross-fitted bigram model and a feature warden over length, punctuation, digits and domain-term counts.

For detailed attribution, we add a control group with the same renderer, the same 20 surface facts, no second stream and no keyed label. These plain sentences already seperate from real news at 0.808. On the same cover the carriers reach 0.952. It shows that the most of 0.952 is the renderer's own distribution, short LLM-generated text vs news. The second reading contributes 0.14. Against Wikipedia the result is 0.888. We also built two covers whose topic matches the carriers' domain, technical Wikipedia articles and technology news.The carriers separate from them at about 0.98.

This is not a tuning problem. We tried five ways to reduce the 0.14: banning the domain terms, resulted in unencodable frame, for capacity failure at position 3; allowing only two generic terms (0.947); density-penalised fine-tuning; LoRA style alignment (0.929); and a warden-in-the-loop round (0.916). The best of them is still far from our own target of 0.85.

Our understanding is that the residual is not the key but the renderer: under the current setting, a renderer can carry the second reading only through explicit domain vocabulary, and the vocabulary, however, is exactly what a feature warden finds. In an on-going follow-up work, we observe that even when a model is given the relevant knowledge, it does not shift its reading the way human readers do on structure-level cues (although it does on fact-level cues). If that holds, the gap will be closed by renderers that express a second reading implicitly, rather than by a bigger knowledge base.

Critical Weakness 3: no semantic-pivot warden

Thank you for pointing out it. This question is partly answered in the submitted manuscript, and we ran a measurement during the rebuttal period as a try to answer the other part.

The paper already evaluates a warden aimed at the construction(§4.5 and Appendix A.5). During the rebuttal, we added another evaluation to determine whether a coordinate can be extracted from a text. With a public rule extractor, extraction succeeds on 69.71% of the carriers and on none of the news sentences or the Wikipedia ones.

The result shows that the pivot trace is visible when the carriers are put next to natural text. But it does not mean the key is leaked. As aforementioned, it is the domain-cue signal. On the coordinate distribution, the key is at chance(0.511) conditional on extraction.

We do not expect a better choice of domains can improve the result, because the attribution is structual. In the protocol settings, the arrier has to be extractable and ordinary text has to be not. Under the same extractor, any gain on one side is paid on the other. Since the carriers keep at least one domain term per sentence by construction, while natural text mentions such terms only occasionally, that density difference is enough to separate these two at 0.988. What changes the balance is a different carrier rather than a larger vocabulary space. Cues carried by how the sentence is built is more promising than by which nouns it contains. That is what we think the next construction must do.

AuthorsOpenReview ↗

Response to Reviewer Mq4w - Part 3

Critical Weakness 4: manually constructed semantic graph and generalization

We do agree that a single hand-built resource cannot establish generality. We would have chosen any existing resource, if possible. Unfortunately, to our best knowledge, none of the ones fits the project. What we need is not a word with two senses. It should be a pair of checkable facts, one in each domain, together with a public context that licenses both, and a set of cues by which a receiver holding the hypothesis can extract the second reading from the sentence.

We have examined as many possible candidates as we can. Pun corpora seemed to be the closest, but we finally found them quite different. ChinesePun (2.1k puns, Chen et al., ACL 2024 Findings) was built for pun generation: an entry is a sentence that exploits one word with two readings, and the annotation is about the pun. It supplies pivots, not payloads, that means no paired facts, no public context, no cue that a receiver can extract. SemEval-2017 Task 7 is English corpus with detection, location, interpretation, but supplies no paired facts either. Lexical resources such as WordNet or HowNet give senses and sense relations, but the cross-domain mapping between two ontologies is precisely the part that has to be judged case by case.

Ultimately, as a practical choice, we presented the semantic graph in the submission. Scaling it is a curation problem rather than a scraping problem. Every added unit has to be checked for realizability and for extractability, which is why the resource is small. Scaling would improve coverage with more domains and more contexts, but not stealth. The reason is given in the previous answer: the carrier has to stay extractable while ordinary text does not. The direction we are taking is to make the construction itself cheaper, rather than to scrape more text. And semi-automatic curation plus per-unit checking may work.

Answer to Question 1

The two operating points below are both end-to-end and both count framing, authentication and repetition; both are from the pivot recipe, with 10/10 frames encoded and recovered and zero erasures.

payload per streamphysical positionspayload bitsR_net (bit/carrier)calls per encoded framecalls per payload bit
1 B240160.0673,142196
8 B4081280.3145,34342

The two rows differ by 4.7× in rate and in cost per payload bit. The 240-position frame nominally carries 480 label bits and delivers 16, so 0.067 is an operating point we chose, not the ceiling of the channel. The ceiling is the label width: acceptance falls as 2^-q, and retry exhaustion is 0.9–1.8% at q = 3 and 85.9–94.8% at q = 4.

One number needs a note: the submitted run in the manuscript used 4,407 calls per frame at the 1 B point, while the run reported during this discussion uses 3,142. The difference is not in the accounting. We use the same configuration. But the number of rejection-sampling attempts per position varied (6.82 at submission, 4.04 now). One possible reason is that the Deepseek service behind the APIs has been updated between the submission and the discussion.

Answer to Question 2

This is a good place to state the motivation behind Iphis. First of all, the difference between one stream and two is not speed, it is who can read what.

As aforementioned, the inner meaning can only be decoded with both key and knowledge base. That is to say, what we pursue is not an encrypted sentence, but a text with hidden intent. The second reading can only extracted by reader with proper knowledge. Just as the example we offered in §1: "A container flies up to the sky, and the moon shines over Chang'an." This sentence is undoubtedly weird, but nursery rhymes cotaining something weird are often considered to be acceptible. People ignored its spread, and the emperor extracted the implications. He did hold the mapping: "container, moon" → the general's name, "flying up to the sky, shining over Chang'an" → threat to the emperor(Similar ambiguous mappings are also very common in Nostredame's "Les Prophéties"). It was a special natural case of information hiding. The genre absorbs the weirdness, and the knowledge decides who can see it. On the other hand, however, this rhyme is merely a plain piece of classical literature, and has no ability to resist unauthenticated extraction. So we choose a keyed second stream, not a plain one.

This is also the honest boundary of our present construction. The domain vocabulary appearing in sentences is the weirdness and cause the separability we reported before. We are looking for proper genres that can absorb the weirdness, which are not implemented yet.

Comments on the paper

1 comment
Area_Chair_Z841OpenReview ↗

Please Participate in the Discussion Phase

Dear Reviewers,

The authors have posted their rebuttal and responses addressing your review comments and questions.

Please carefully read their responses (as well as the updated manuscript, if applicable) and actively engage in the reviewer-author discussion phase. If you have any remaining concerns or follow-up questions, please post them so the authors can address them before the discussion period closes.

Your timely participation and updated assessments are essential to helping us reach a fair and well-informed recommendation for this submission.

Best regards,

Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.