acceptodds
Under review as a conference paper at ICLR 2027

The Right Source Is Not Enough: Understanding History Interference in Speech-LLM Memory

Abstract

Speech large language models (Speech-LLMs) rely on audio history to support multi-turn conversations. Existing work on conversational memory has largely examined whether past information is retained and can be retrieved when needed. We study a different failure mode in which irrelevant historical audio changes an answer even when the relevant event remains identifiable. Using controlled multi-source histories and causal interventions on two Speech-LLMs, we find that irrelevant history systematically shifts predictions toward values mentioned in another event. Internal interventions show that the queried event remains represented and causally effective, ruling out source identification as the main explanation. The causal path further depends on answer format. In open-ended QA, history-value errors largely disappear once later computation can no longer access the irrelevant audio, whereas many multiple-choice errors persist because historical influence has already entered question-candidate states. Layer-wise interventions reveal a non-serial decision process. Sensitivity to answer-value information is strongest in intermediate layers, while sensitivity to source role becomes stronger deeper in the network. On natural failure cases, strengthening the correct source role and neutralizing the competing historical value can each recover errors, with their joint intervention producing the strongest recovery. These results identify historical interference as competition between source-role and answer-value influences during downstream decision making, rather than a general long-context degradation or a failure of source identification alone.

Primary area:
interpretability and explainable AI
Keywords:
audio language models, multi-turn audio, conversational memory, historical interference, mechanistic interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ADGbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how irrelevant historical speech affects Speech-LLM answers. Through causal interventions on two models, the authors show that historical interference can persist through intermediate hidden states, even after direct access to the source is blocked. They further identify distinct source-role and answer-value influences, suggesting that correct source identification alone is insufficient to prevent memory interference.

Critical strengths

The work focuses on an important challenge in Speech-LLM memory. The authors provide controlled experiments across two models, offering evidence of historical interference. The combination of attention masking, state transfer, and subspace interventions provides causal insights beyond behavioral evaluation.

Critical weaknesses

My main concern lies in the experimental design and interpretation. The authors do not clearly explain why the role subspace is constructed before the value subspace, as reversing this order may lead to different results. The role subspace is also derived from hidden-state differences between different questions, but these differences may reflect variations in question wording, semantics, or answer content rather than source-role information alone. Additionally, the role intervention changes the model's answer preferences, but this does not necessarily mean that the model has correctly identified the relevant source. Since the role subspace is derived from hidden-state differences between questions. Without directly verifying whether the model identifies the correct source, it remains unclear whether historical interference persists despite correct source identification.

Questions for authors

Could you test whether reversing the construction order of the role and value subspaces changes the results?

Could you use different question templates and paraphrases to verify that the role subspace captures source-role information rather than question wording or answer content?

1 comment

Thank you for the specific questions. We tested both construction orders and frozen-basis wording controls, and measured source associations independently. Question/content contrasts below correspond to the original role/value terminology. References are to the revised manuscript.

1. Why question-first, and what changes on reversal?

Question-first asks whether historical content outside the fitted question directions still affects the answer: the joint-minus-role-only test in original Section 3.3, Eq.(3). We fit question differences with history fixed, remove their projection from content differences, then fit the residual. Reversal changes this reference comparison, not an assumed processing order (Appendix D.1, Eq.10, p.27).

The subspace materials contrast an original arrangement and its update; the receiving question asks about the original. Both orders were evaluated at all 36 layers on 512 test scenarios, with bases fitted on 128 development scenarios. At layer 23, question-only mean shifts under question-first/content-first are 2.154/1.993 logits in MOSS and 0.811/0.234 in MiMo (Table 17, p.30).

Why do effects change? Let A/B denote question-first/content-first, PO=PQO+PCOP^{O}=P_{Q}^{O}+P_{C}^{O}, and dQ,dCd_{Q},d_{C} the fixed donor contrasts. The joint edit is eO=PQOdQ+PCOdCe_{O}=P_{Q}^{O}d_{Q}+P_{C}^{O}d_{C}. Eq.(11), p.28 gives, before rounding:

eB−eA=(PB−PA)dC+(PQB−PQA)(dQ−dC)e_{B}-e_{A}=(P^{B}-P^{A})d_{C}+(P_{Q}^{B}-P_{Q}^{A})(d_{Q}-d_{C}).

The first term reflects a change in the retained joint span. The second reflects which donor contrast supplies coefficients within it. Even an unchanged joint span need not preserve the edit. MOSS's joint-span overlap is 0.984, yet its joint margin shift changes from 2.719 to 2.663. Question-only edits instead differ by (PQB−PQA)dQ(P_{Q}^{B}-P_{Q}^{A})d_{Q}. After subsequent computation, output changes depend on the final-state difference along the answer-readout direction (Eq.12). These tests retain projected magnitudes, so direction and amplitude both change; the output effects are not separately attributed to the two terms above.

The shared finding is additional content influence. At layer 23, joint-minus-question-only effects are 0.565/0.670 logits in MOSS and 0.405/0.511 in MiMo. This mean effect favors the unasked answer at all five prespecified reporting layers, in both models and orders, with every pointwise interval above zero (Table 16, p.29; complete-audio-group resampling).

2. Do the directions reflect wording or source role?

With bases frozen from the original templates, we test receiver paraphrases, donor paraphrases, and same-event wording differences that preserve the requested event and answer. Tests cover two paraphrases per task family, all 512 scenarios, both orders, and five layers (Appendix D.2-D.3, Tables 18-20, pp.30-33).

At layer 23, question-first receiver effects for original/p1/p2 wording are 2.154/2.117/1.969 logits in MOSS and 0.811/0.726/0.813 in MiMo. Same-event rewording also contributes: the MOSS original-event p2 control gives 0.114 [0.079, 0.150] logits.

The effect therefore transfers across tested phrasings, but paraphrases preserve both event and answer, whereas switching events can change both. We use question/content contrasts operationally, not as pure semantic variables. Source associations are tested independently.

3. What verifies source associations on original errors?

In Section 4.1 (Figure 3 and Table 4, pp.5-6), three linear probes read layer-17 states: one identifies the requested utterance, and two identify each utterance's associated option. They are fitted on development data and frozen. The model answers without their predictions.

All three are correct on 488/617 MOSS and 206/218 MiMo distractor errors (79.09%/94.50%). Swapping utterance order while fixing questions/options reverses source assignments. All three probes remain correct in both orders on 85.23%/97.46% of 4,096 pairs; question-to-source random-label controls are near 50% (Appendix A.3, pp.15-17). An unchanged question-to-position mapping cannot satisfy both orders.

Among correct/error pairs where all three probes are correct in both unmodified runs, transferring wrong-run option Keys and Values to the correct partner reproduces the specific wrong answer in 190/215 MOSS and 83/118 MiMo pairs (Section 4.4; Table 13, p.22). Only the distractor differs; question/options stay fixed. Transfers act on later reads of option vectors, not entire answer states.

Thus, the associations needed for a correct answer remain recoverable while later option reads can favor the distractor. This does not assert that the decoder already selected the correct source, and does not depend on pure role/value separation.

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

The paper studies errors caused by irrelevant speech history in MOSS-Audio and MiMo-Audio. Natural recordings and synthetic experiments show value-directed interference. Source blocking and bidirectional state transfers distinguish two patterns. Open-ended errors largely depend on continued source access; multiple-choice errors can persist through question states.

Critical strengths
  1. Under fixed source blocking, bidirectional state transfers suppress and restore history-driven errors. This supports a specific error-propagation pathway.
  2. Matched histories, answer-format controls, and resampling by episode strengthen the evidence.
Critical weaknesses
  1. The work is primarily analytical. How its findings can guide practical training and inference remains unclear.
Questions for authors
  1. How can this analysis guide improvements in the user experience of interacting with the model?
1 comment

Thank you for asking how the analysis can improve interactions with the model. We have revised Section 4.5 to make the practical implications explicit and tested two analysis-guided improvements in QA. The training experiment supervises correct answer updates and stability. The frozen-backbone adapter adjusts how retained audio contributes to the current answer. Both improve QA accuracy in the reported evaluations (Table 8, p.10). References below are to the revised manuscript.

1. Train both correct updating and correct stability

Source associations remain recoverable on many errors, and later option-read interventions can correct answers while both utterances remain available (Sections 4.1/4.4). This motivates supervising how retained information contributes to the current answer.

Consider two histories: the suitcase stays green, while the backpack is white in one and yellow in the other. Asking about both objects in both histories gives four inputs. The backpack answer should change; the suitcase answer should remain green.

We use ordinary answer cross-entropy and rank-8 LoRA in attention and FFN projections. Training uses 128 scenarios; evaluation uses 8,192 inputs from 512 scenarios with no shared audio, but the same templates and voices. Supervised results are means over three seeds (Section 4.5, Table 8a, pp.9-10):

ModelAccuracy (%)Correct update (%)Correct stability (%)Distractor answers (%)
MOSS84.12 -> 96.4676.17 -> 94.9277.98 -> 95.0215.72 -> 3.30
MiMo95.31 -> 99.7693.26 -> 99.5491.50 -> 99.674.36 -> 0.02

Correct updating and stability require both paired answers to be correct; repeating a wrong answer is not stability. The design gives reusable examples and measures for selective history use. It tests alternative histories, not successive spoken corrections, and does not isolate pairing from equally sized unpaired training (Appendix C.4, pp.25-26).

2. Adjust audio contributions during question processing

Equation (2), Section 4.5 (p.9), shows that attention-weight changes affect the answer through the Values and subsequent computation. An attention weight alone does not determine the required adjustment. We test the broader implication of adjusting audio contributions.

With the backbone frozen, a rank-32 adapter uses the current attention Query, each audio turn's weighted Value contribution, and their sum to correct the head output during the question turn. It preserves ordinary attention weights at each adjusted read and does not remove clips. Both learned evaluations use no source cue, gold answer, or donor activations at inference.

On 517 MMAR-derived questions, each tested with two, three, and four clips (1,551 responses), Table 8b (p.10) reports:

  • MOSS: 54.48% -> 59.12%; +4.64 percentage points, 95% CI [1.80, 7.42].
  • MiMo: 53.32% -> 59.45%; +6.13 points, 95% CI [3.20, 9.16].

Test parent videos are disjoint from training. Each model's adapter was trained once; intervals resample 115 complete parent-video groups with all questions and clip counts retained, measuring test-sampling uncertainty rather than training-seed variation (Appendix C.5, pp.26-27). This is our multi-clip adaptation of MMAR, not its official protocol.

These results establish improved multi-clip QA. Single-clip accuracy also improves, so the gain cannot be attributed specifically to interference suppression. The adapter tests an implication of the analysis, not a deployable version of oracle option-Key transfer.

3. When a reliable source restriction is available, apply it during question processing

The state-transfer tests show that current question/option states can retain distractor influence after later direct access is blocked (Section 3.2, Table 3, p.5). We test the implication using the correct source index as an oracle control.

With identical source-instruction prompts, blocking only at the answer position yields 93.26%/98.74% accuracy in MOSS/MiMo. Blocking throughout the current question turn yields 99.37%/99.15%, gains of 6.11/0.41 percentage points (Appendix C.4, Table 15, p.26). Both conditions block at all 36 layers; only the restricted token positions differ.

This gives conditional guidance on when to apply an available source restriction; it does not evaluate how to select that source.

Benefit for interaction

A user should be able to ask about a retained event without receiving another event's answer. The analysis yields a tested training design, measures of correct updating and stability, and an internal adjustment that improves multi-clip QA.

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

This paper addresses the well-known problem of hallucination from irrelevant context in speech language models through behavioral analyses and interventions. It distinguishes role and value directions in hidden states and shows that manipulating these directions can alter answer selection and correct some history-induced errors.

Critical strengths
  1. The paper addresses the important issue of hallucination caused by irrelevant context in speech language models.
  2. The experiments show that these errors can be investigated through targeted interventions on attention and hidden states.
Critical weaknesses
  1. The paper’s analyses are largely trivial and offer a limited contribution. The analyses in Section 4.1 (prediction bias toward candidate answers appearing in irrelevant history), Section 4.2 (prediction shifts through direct or PCA-based latent-state transfers), and Section 4.3 (attention manipulation and state transfers across access conditions) all yield unsurprising results. What these intervention-induced prediction changes establish about the original errors remains unclear. The paper can largely be read without Sections 4.1–4.3, as their findings are not clearly integrated into the analyses in Sections 4.4–4.5. In Section 4.4, the proposed directions are not shown to disentangle role and value information (changing the queried event may also alter answer-related representations, rather than isolating role information alone), making the analyses attributing their intervention effects to distinct role and value influences questionable. Section 4.5’s additional explanatory value beyond the controllability demonstrated in Section 4.4 is unclear, especially because its intervention shifts use oracle knowledge of the test set. This provides a narrower demonstration of example-specific correction rather than stronger evidence for the proposed explanation of historical interference.
  2. The paper’s writing needs substantial improvement. Despite unused space in the main text, essential experimental details are deferred to the appendix, making the analyses difficult to understand without repeatedly consulting it. Each experiment would benefit from a concrete worked example or a figure clearly identifying the inputs, transferred states, and measured outcomes. The terminology is also confusing and inconsistent. In particular, “history” sometimes denotes the entire preceding audio context, including the target, and elsewhere denotes only the distractor source contrasted with the target. Consistently using “relevant source” and “distractor source” when distinguishing their roles would make the experimental setup and results clearer.
Questions for authors

See weaknesses. Please let me know if there is any misunderstanding of the paper. Also, Figure 2 explicitly refers to source B, whereas the experimental questions identify events or objects rather than source indices. My understanding follows the experimental description, but please let me know if otherwise.

1 comment

Thank you for the detailed review. We have substantially revised the manuscript, clarified how the analyses build on one another, standardized terminology throughout, and made the key comparisons explicit in the main text. The revision connects a defined option-reading computation to specific errors made by the unmodified models. References below are to the revised manuscript.

1. What do interventions establish about original errors?

In Section 4.4, the relevant utterance, question, and options stay fixed; only the distractor changes. For example, a backpack color change can corrupt the suitcase answer (Table 1, p.2).

In layers 18-35, we replace option Keys/Values only when positions after the options read them. We do not copy full option hidden states or final-answer states. Correct-partner Keys repair 269/273 MOSS and 113/123 MiMo errors. In reverse, wrong-partner Keys and Values reproduce the exact original wrong choice in 240/273 and 87/123 pairs (Table 7 and Figure 6, p.8).

Repair alone need not reproduce an error: the correct answer's mean lead exceeds the distractor's in the paired runs (Table 13, p.22). The reverse test adds evidence beyond correction.

2. How do the preceding analyses support this explanation?

Section 3.1 establishes bias toward the particular content in the distractor, with fixed-length controls (Figure 2, p.4). Under the same later-source block, bidirectional question/option-state transfers suppress and restore the original distractor choices (Section 3.2, Table 3, p.5). This motivates tracing how current text states influence later answer computation.

Section 4.2 tests a downstream dependency of the early FFN intervention. Keeping patched gates active, restoring the receiver's option Keys leaves 12%/20% of the mean induced score shift in MOSS/MiMo; restoring Values leaves 89%/86% (Figure 4, p.7).

Section 4.3 then isolates the read without patching gates or option states. Redistributing attention between and within options nearly matches the mean directional Key-transfer effect. Within each intervened read, Values, total option attention, and non-option weights stay fixed (Figure 5, p.7). Later Queries can still evolve.

These component tests retain both utterances, unlike the source-blocking protocol. They identify how later tokens weight already-computed option content. Section 4.4 connects those same reading connections to the model's particular original errors.

3. Does this require pure role/value subspaces?

We tested both fitting orders and frozen-basis paraphrases, including same-event wording controls (Appendix D, Figure 9 and Tables 16-20, pp.27-33). Effect sizes depend on construction. Mean additional content effects remain positive at all five prespecified layers in both orders and models, but changing the queried event can also change answer representations. We use question-contrast/content-contrast rather than claim pure semantic separation.

The revised explanation is independent of that decomposition. Three frozen probes recover the requested utterance and both utterance-to-option associations on 79.09%/94.50% of MOSS/MiMo distractor errors (Section 4.1, Figure 3 and Table 4, pp.5-6). This measures recoverability, not a completed native source choice. When all three probes are correct in both original runs, joint Key/Value transfer still reproduces the specific error in 190/215 and 83/118 pairs (Table 13, p.22). The FFN/attention tests use computationally defined components, not fitted contrast subspaces.

4. What is learned despite using oracle donors?

The donors are diagnostic counterfactuals, not an inference-time solution. Figures 4-5 cover all 4,096 changed-distractor directed inputs per model, without selecting by original correctness. The layer split was not optimized on test errors (Appendix A.2, p.15). Table 7 separately selects correct/error pairs from unmodified predictions, not intervention success.

The evidence combines directional score changes, downstream restoration, fixed-Value attention redistribution, and exact-error reproduction. These test which computations transmit a specified distractor change. Separate learned evaluations improve QA without source cues, gold answers, or donor activations at inference (Section 4.5, Table 8, pp.9-10).

5. Presentation and source B

Your reading is correct: the main questions identify an event or object, not a source index. The original Figure 2 wording was misleading. Tables 1-2 (pp.2-3) and Figure 6 (p.8) show the actual questions; the explicit-source oracle is separate (Appendix C.4, Table 15, p.26).

We now distinguish dialogue history, relevant utterance, and distractor utterance. Section 2 defines donors, receivers, transfers, and metrics; Section 3.2 specifies the copied span and blocking conditions. The revised sequence connects retained associations, biased option reads, and original mistakes.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Main Findings

We thank the reviewers and Area Chair. Our analysis links recoverable source associations to later option reads that influence specific errors. We test implications for history use. References are to the revised manuscript.

1. Connection to original errors

In Section 4.4, only the distractor utterance changes; the relevant utterance, question, and options stay fixed. Correct/error pairs are selected from unmodified predictions, not intervention success. In layers 18-35, we replace option Keys/Values only when positions after the options read them, without copying full option or final-answer states. Correct-partner Keys repair 269/273 MOSS and 113/123 MiMo errors. In reverse, wrong-partner Keys and Values jointly reproduce the exact original wrong choice in 240/273 and 87/123 pairs (Table 7 and Figure 6, p.8).

Three frozen probes independently recover the requested utterance and each utterance's associated option (Section 4.1, pp.5-6). When all three probes are correct in both original runs, joint transfer still reproduces errors in 190/215 and 83/118 pairs (Table 13, p.22). These probes measure available associations, not a completed native source-selection step.

2. Computational dependencies

Section 3 shows content-directed interference and its persistence in question/option states under the tested later-source blocks. This motivates tracing subsequent option reads. The component tests retain both utterances and cover all 4,096 changed-distractor directed inputs per model, without selecting by correctness.

With early FFN gates patched, restoring receiver option Keys leaves 12%/20% of the mean gate-induced score shift in MOSS/MiMo; restoring Values leaves 89%/86% (Section 4.2, Figure 4, p.7). Separately, redistributing attention between and within options nearly matches the mean directional Key-transfer effect. Values, total option attention, and non-option weights remain fixed within each intervened read (Section 4.3, Figure 5, p.7).

The same interventions produce larger mean shifts toward the donor content when it is asked about, but positive shifts remain when it is unasked (Table 6, p.7). Thus, the question modulates content influence without fully restricting it to the requested utterance. Together with Section 4.4, these tests connect early gate-induced changes, later option reads, and specific original errors.

3. Subspace construction and wording

Question-first construction tests additional content influence outside fitted question-contrast directions; it does not assume an internal processing order. We tested both orders at all 36 layers. Reversal can change the retained span and which donor contrast supplies the edit coefficients (Appendix D.1, Eq.11, p.28). Effect sizes change, but the mean additional content effect favors the unasked answer at all five prespecified reporting layers in both models and orders, with pointwise intervals above zero (Table 16, p.29).

Frozen-basis paraphrases show transfer across tested phrasings; same-event controls also reveal wording contributions (Tables 18-20, pp.31-33). We therefore use question-contrast/content-contrast for the original role/value directions, without claiming pure semantic separation. The probes and component tests do not use these fitted subspaces.

4. Tested practical implications

Answer supervision on relevant-change and irrelevant-change pairs improves correct updating and correct stability, each requiring both paired answers to be correct. Across three seeds, distractor-answer rates fall from 15.72% to 3.30% in MOSS and 4.36% to 0.02% in MiMo on disjoint test audio with shared templates and voices (Section 4.5, Table 8a, pp.9-10). This tests the training design as a whole.

A frozen-backbone adapter adjusts audio contributions during question processing. On 517 MMAR-derived questions with two to four clips, accuracy improves by 4.64/6.13 percentage points in MOSS/MiMo (Table 8b, p.10). Single-clip gains prevent attributing this improvement specifically to interference suppression. Both learned tests require no source cues, gold answers, or donor activations at inference. A separate oracle control gives conditional guidance: when the correct source is known, restricting access throughout the question turn works better than restricting only the answer position (Table 15, p.26).

We standardized dialogue history, relevant utterance, and distractor utterance, and clarified the comparisons in the main text. Tables 1-2 and Figure 6 show actual event/object questions, correcting the earlier source-B schematic. The individual replies give further details. We ask the reviewers and AC to consider this connected account of incorrect history use and its tested practical implications.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.