acceptodds
Under review as a conference paper at ICLR 2027

Characterizing the Gap Between Knowing and Answering in Speech Language Models

Abstract

A large speech language model can learn a fact from text and answer a written question correctly, but still fail to answer the same question through speech. This discrepancy does not establish whether the learned information is absent from speech processing, is present but not affecting the response, or contributes to a response that still fails to recover the answer. We investigate these possibilities by teaching models new facts through text-only fine-tuning and examining how this learning changes their hidden states when processing speech. By changing the learned answers and selectively removing these changes, we test whether they reflect answer content and influence generation. Our results show that information about the learned answer can remain present and affect the model's response even when speech recall fails. On the speech failures we examine, removing the changes associated with the queried fact harms responses more than removing equally strong changes associated with other facts. We observe this content-sensitive effect across two model families and under changes in question wording and synthesized voice. However, in a controlled ablation study, increasing the measured changes beyond their original strength provides little additional improvement in average answer quality. These findings show that a model can use knowledge learned from text and still answer a spoken question incorrectly. Evidence that this knowledge contributes to a response does not establish that simply strengthening its contribution will produce the correct answer.

Primary area:
interpretability and explainable AI
Keywords:
speech language models, cross-modal knowledge transfer, factual recall, mechanistic interpretability, causal interventions

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xbBPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper talks about this specific modality gap question, that usually arises in Speech Language models, where a knowledge which model gains in Text, sometimes is missed in the Audio modality. Here the paper takes an Interpretability based investigation into providing some insight into this problem. Two possible interpretations are,

  • The text and speech representations are completely separate. within the model
  • The magnitude of contribution from text is not enough for the Audio modality pathways.

The paper picks an intermediate layer output interventions on two models Qwen2.5 Omni7B and MiMo Audio 7B Instruct, fine tuned on TOFU QA pair dataset, to see its effects. The paper identifies the following notions to do the analysis.

  • A speech update vector as the hidden state difference between text-fine tuned model and base model difference on speech inputs.

Then it tries to see the effects on removal controlled for by other parameters.

  • Does speech update contribute to answering? By comparing the removal of same-question vector VS matched control vector removal.
  • Does speech update vector contain traces of learned answer?: Controls for effect of the answer part alone, by using Vector effect differences from a True Answer model VS Counterfactual answer model.
  • Amplification effects of this Vector and its effects.

The results seem to show, that the answer information is present even when model is unable to answer correctly.

  • The vector has answer information traces, not just the question. Removing own vector hurts performance than other model's vector.
  • The effect shows up even when speech recall fails: This means there is still textual knowledge information in the vector, but model is unable to make use of it.
  • Amplification does not seem to help
  • Swapping Audio state vector with Text state vector seems to help
Critical strengths
  1. The paper deals with a very clear and actionable problem to tackle. It poses in a controlled setting, and probes it in a well defined way.
  2. The experiments are all well controlled for. Such as
  • the text-training and speech-inference,
  • the fictional authors rule out any pretraining concerns etc.
  • The results also control for marginal effects instead of base effects. such as comparing same question removal to other vector removal
  1. All the different variations of the investigation all point to the same verdict that textual knowledge is present within the model in speech mode.
  2. Running the experiments against two model families with multiple variations makes the results much more trustable and robust.
  3. A good amount of peripheral supporting information in the Appendix.
Critical weaknesses
  1. The two models, one dataset, one voice testing setup might pose this in a somewhat toy setting without real world effect on this.
  2. The heavy textual fine tuning representations and studies may not translate well to pretraining setup.
  3. Finally the results, directionally tell the correct story, but in magnitude the effects look quite small. Not sure how generalizable they are.
  4. Not a lot of motivation behind picking this one vector from one layer. No studies on what motivated this.
  5. Retrieval accuracy in Qwen's final layer drops. But paper doesnt talk much about it.
  6. The paper does not raise a conversation towards how to fix the modality, but just answers that the text knowledge traces exist in speech mode also.
Questions for authors
  1. Any studies done towards choosing layers for intervention in the models? How did you choose this? Was there a better choice for this? Any sweep studies done for this?
  2. Since the models are trained differently the vector comparison could be very much affected by the training dynamics and pathway it took. Any studies on the effect of this? Anyway to get better comparability between these two vector spaces?
  3. More importantly, what intuition does this give you in closing the modality in Speech Language Models? Either in training setting or in data mixture settings?
  4. Didnt quite understand the significance of Figure 2a. My understanding is that, it simply shows that may be the question information is represented in the hidden states? Is there more to it?
  5. What might be other factors that might explain the modality gap? anything that came up in your studies?
Reviewer_xDGNOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies why a speech language model may correctly answer a fact learned through text when queried with text, yet fail to answer the same question when queried through speech. The authors introduce a controlled text-write, speech-query setting and define a speech-update vector as the change in a speech hidden state induced by text-only fine-tuning. They use causal interventions to examine whether this update reflects the learned answer and whether it contributes to generation when speech recall fails.

The results show that the update contains answer-specific information and has a measurable causal effect on the response: removing the update associated with the queried fact harms answer quality more than removing norm-matched control updates, including on examples that are already answered incorrectly from speech. However, amplifying the same update beyond its original strength provides little additional benefit, while replacing the speech hidden state with the corresponding successful text state improves recall. The main findings are validated on Qwen2.5-Omni-7B and MiMo-Audio-7B-Instruct, with additional controls involving counterfactual answers, alternative question wording, and synthesized voices. Overall, the paper provides evidence that failed speech recall does not necessarily imply that text-learned knowledge is absent from, or unused by, the speech computation.

Critical strengths
  1. The paper makes a clear and meaningful causal distinction between information being represented and information being used. A major strength is that the analysis goes beyond probing whether learned information can be decoded from speech hidden states. By removing question-specific speech-update vectors and comparing them with norm-matched controls, the authors show that the learned update has a direct behavioral effect on generation, even for examples where the model ultimately answers incorrectly. The counterfactual-answer experiments further support that these effects are related to the learned answer rather than only to the identity of the question. This provides stronger evidence for the paper’s central claim than representation similarity alone.

  2. The experimental analysis is unusually careful and well controlled for a mechanistic study of the speech-text gap. The paper evaluates the main finding through several complementary tests, including vector removal, answer-preference scoring, frozen-head readout, cross-modal update retrieval, vector amplification, and text-state replacement. It also reproduces the main removal effect with a second model family and under changes in question wording and synthesized voice. These controls make the central observation—that text-learned information can contribute to a speech response without being sufficient for correct recall—substantially more convincing.

Critical weaknesses

1. The central claim is supported mainly in a direct QA memorization setting, which substantially limits its interpretation as general text-to-speech knowledge transfer.
The main experiments fine-tune the models on 4,000 question–answer pairs and then evaluate recall on the same questions presented through text or synthesized speech. As a result, successful text recall may largely reflect memorization of a directly supervised question-to-answer mapping, rather than acquisition of a fact that must later be retrieved under a novel query. This distinction is important because the paper frames its conclusions broadly in terms of knowledge learned from text. The factual-statement experiment in Appendix F.3 is closer to this broader setting, but the resulting effects are much weaker and depend on question wording. Therefore, the current evidence does not establish whether the main causal findings extend beyond directly supervised QA mappings. This limits the generality and significance of the paper’s central “knowing vs. answering” interpretation.

2. The speech-update vector is a broad fine-tuning-induced state difference, so its interpretation as a factual-knowledge signal remains only partially established.
The speech-update vector is defined as the full difference between the hidden states of the Text-FT and Base models. The paper provides useful controls showing that this difference contains answer-specific information and that removing it has question-specific behavioral effects. However, the vector may also contain other changes introduced by fine-tuning, such as changes in task behavior, confidence, output preferences, or other aspects of the representation. The current experiments therefore support the claim that an answer-related fine-tuning signal causally contributes to the response, but they do not cleanly isolate the factual component responsible for this effect. This matters because the paper’s main interpretation relies on treating the observed intervention effect as evidence that the model is already “using” the learned knowledge during failed speech recall.

Questions for authors
  1. Do the main causal findings persist when the evaluation question is not seen during fine-tuning? The main experiments train and evaluate on the same QA pairs, while the factual-statement experiment in Appendix F.3 appears substantially weaker and sensitive to question wording. Can the authors clarify whether the key findings—particularly answer-specific speech-update effects and the stronger damage from same-question removal—also hold when a fact is learned from declarative text and later queried using previously unseen or paraphrased spoken questions? If not, would the authors agree that the current conclusions should be restricted to cross-modal recall of directly supervised QA mappings rather than text-learned knowledge more generally?

  2. How specifically can the behavioral effect of the speech-update vector be attributed to factual information rather than other fine-tuning-induced changes? The counterfactual-answer and matched-control experiments show that the vector contains answer-specific information, but the vector itself is still the full hidden-state difference between the Base and Text-FT models. Can the authors provide additional evidence that the component responsible for the observed behavioral effect is specifically tied to the learned factual content, rather than to other correlated changes introduced by fine-tuning? Clarifying this point would materially affect how strongly the intervention results support the claim that learned knowledge is already being used during failed speech recall.

Reviewer_g9QqOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates why speech language models may fail to answer spoken questions even after successfully learning the relevant facts through text-based fine-tuning. The authors distinguish between the absence of learned knowledge from speech processing and the presence of knowledge that influences generation but remains insufficient for correct recall. The paper introduces a speech-update vector, defined as the difference between the hidden states of a text-fine-tuned model and its base model when processing the same speech input. Using this representation, the authors conduct controlled interventions to examine whether the learned answer is encoded in the update, whether the update causally contributes to the model's response, and whether amplifying it improves answer quality. Experiments on Qwen2.5-Omni-7B and MiMo-Audio-7B-Instruct use 4,000 question-answer pairs about fictional authors from TOFU, with spoken questions generated using text-to-speech. The authors compare same-question vector removal against magnitude-matched control interventions, construct counterfactual fine-tuning conditions to test answer specificity, and investigate vector amplification and cross-modal hidden-state replacement. The results suggest that text-learned information can remain present and causally influence speech responses even when the model answers incorrectly. Removing the question-specific update degrades answer quality more than removing matched control updates, whereas amplifying the update beyond its original strength provides little additional benefit. Replacing the speech hidden state with the corresponding text hidden state can improve responses, suggesting that the text-to-speech gap involves more than the simple absence of learned factual information. Overall, the paper offers a controlled, intervention-based analysis of the distinction between knowledge representation and successful knowledge retrieval across modalities. Its main contribution is diagnostic rather than directly algorithmic: it provides evidence about why text-learned knowledge may be insufficient for correct speech-based answering, while leaving the broader mechanisms responsible for the modality gap unresolved.

Critical strengths
  1. A well-motivated research question with a carefully designed causal analysis. The paper addresses an important question that cannot be answered through output accuracy alone: whether knowledge acquired through text continues to influence a speech language model's response when spoken recall fails. The speech-update vector provides a concrete way to investigate this question. Particularly valuable are the magnitude-matched control interventions and counterfactual fine-tuning experiments, which help distinguish question-specific, answer-related effects from generic changes induced by fine-tuning. This design makes the central findings more convincing than evidence based solely on hidden-state probing or cross-modal performance comparisons.
  2. Multiple complementary experiments support the main conclusions. The authors examine vector removal, answer preference, vector amplification, and hidden-state replacement using two speech language model families. They also evaluate alternative question wording and a second synthesized voice, with author-level bootstrap analyses and leave-one-author-out checks providing additional robustness evidence. Importantly, the authors distinguish the causal contribution of a learned update from the ability to produce a correct answer: removing the update can harm a response even when the original answer is wrong, while amplifying the update does not necessarily improve recall. This nuanced interpretation is a meaningful strength of the paper.
Critical weaknesses
  1. The experimental setting limits the generalizability of the conclusions. The experiments focus on two speech language models, fictional factual knowledge from the TOFU dataset, question-answer fine-tuning, and primarily synthesized English speech. Although the additional voice and model-family experiments are useful, they do not establish whether the same mechanisms hold for naturally recorded speech, diverse accents, background noise, or more open-ended factual and reasoning tasks. The results therefore provide strong evidence within a controlled setting, but broader claims about speech language models in general should be made cautiously.
  2. Robustness across training runs and the interpretation of answer quality remain incompletely established. The reported bootstrap analyses quantify uncertainty across questions or authors, but they do not establish variability across independently fine-tuned model checkpoints. Consequently, it remains unclear how sensitive the observed intervention effects are to stochastic variation in training. In addition, ROUGE-L recall is the primary metric for the free-generation intervention experiments. Although the paper includes complementary answer-likelihood and other evaluation analyses, lexical overlap does not always correspond to factual correctness. Reporting exact-match or another semantic correctness measure under the principal removal conditions would strengthen the interpretation of how these interventions affect failed recall.
Questions for authors

1.Were the primary text-fine-tuning, counterfactual fine-tuning, and vector-intervention experiments repeated using independent training seeds? If not, could the authors provide a robustness check across additional independently fine-tuned checkpoints? The current bootstrap procedure addresses variation across sampled questions or authors but does not quantify training-run variability. 2.For the selected speech-failure cohort, can the authors report exact-match accuracy or another semantic correctness measure under the no-intervention, same-question removal, and matched-control removal conditions? This would help establish whether the observed reductions in ROUGE-L recall correspond to changes in factual answer correctness rather than primarily to lexical differences.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.