Pith. sign in

REVIEW 3 major objections 5 minor 56 references

Frozen speech LLMs already hold a recoverable dialogue manager: read residual streams to pick the next conversational move, then steer activations to produce it.

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 ↗

T0 review · grok-4.5

2026-07-30 16:43 UTC pith:X2NCKMXM

load-bearing objection Solid systems paper: internal selection–realization for speech LLMs with real multi-model evidence; classifier-as-metric is the main caveat, not a collapse. the 3 major comments →

arxiv 2607.26928 v1 pith:X2NCKMXM submitted 2026-07-29 cs.CL

Latent-IM: Latent Interaction Management for Speech LLMs

Adar Avsian , Atahan Dokme , Tony Woo , Larry Heck This is my paper
classification cs.CL
keywords speech LLMsdialogue managementactivation steeringconversational movesresidual streamturn-takingtask-oriented dialogue
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.

Classical spoken dialogue systems kept dialogue management separate from wording: a policy chose the next action, and a generator expressed it. Modern speech LLMs bury that choice inside hidden states. This paper asks whether those internal states still encode an analogue of state estimation and action control for everyday moves—acknowledge, check, query, explain, reply—and whether that analogue can be read out and driven without retraining the model. Latent-IM answers yes. A lightweight streaming controller reads residual-stream activations and selects the next move at about 0.60 average accuracy across three task-oriented corpora. Separately learned move-specific activation directions, estimated only on MapTask and transferred unchanged, causally realize a chosen move at generation time, beating strong prompting and decoding baselines. End to end, the system raises match to human moves by 12.5 points over the unsteered backbone and matches LoRA fine-tuning, while a complementary turn-boundary direction grades response length from long to short. The practical claim is that frozen speech-LLM activations are already an observable, steerable interface for local interaction management.

Core claim

An internal analogue of classical dialogue state and action control can be recovered from a frozen speech LLM’s residual stream: a streaming controller selects among five conversational moves from those activations, and move-specific activation steering causally realizes the chosen move at generation time, lifting end-to-end human-move match by 12.5 points over the unsteered backbone while matching supervised fine-tuning—without gold moves or backbone updates at inference.

What carries the argument

Latent-IM: a two-part interface that (1) reads mean-pooled residual-stream activations with a lightweight selective state-space controller to select the next move and a turn length, and (2) injects reusable, Fisher-derived move-specific and turn-boundary directions into chosen layers at generation time, with realization-aware gating so steering is applied only where it lands.

Load-bearing premise

The central accuracy numbers rest on a prompted large-model move classifier being a faithful enough stand-in for human judgments of what conversational move a response performed.

What would settle it

On the same held-out contexts and models, replace automatic move labels with fresh three-annotator majority labels for every generated response; if Latent-IM’s reported gains over the unsteered backbone and over the strongest baselines shrink or reverse under human majority labels, the central end-to-end claim fails.

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

If this is right

  • Dialogue management can be restored as an explicit selection–realization loop inside frozen speech LLMs without fine-tuning the backbone.
  • Move directions learned on one map-style corpus transfer to other task-oriented spoken dialogues without re-estimation.
  • A single turn-boundary direction gives graded, sample-free control over when the model yields the floor.
  • Realization-aware gating (steer only where controllability is high) is required; accurate move prediction alone does not guarantee end-to-end control.
  • The same interface can later select moves by predicted task utility rather than by imitating human move frequencies.

Where Pith is reading between the lines

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

  • If residual streams already carry linearly readable dialogue state, other classical control signals—repair initiation, grounding strength, or floor competition—may be recoverable the same way without new training objectives.
  • Matching human move distributions may be the wrong long-run objective; the paper’s own MapTask mismatch example suggests goal-conditioned selection could outperform imitation on task success even when move accuracy falls.
  • Because steering is applied only at decode time and directions transfer across datasets, the method is a candidate for modular post-hoc policies layered on closed commercial speech models.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper asks whether classical dialogue-management structure—state estimation and action selection/realization—can be recovered from the residual stream of a frozen speech LLM. It introduces Latent-IM: a lightweight streaming controller (selective SSM over mean-pooled activations) that selects among five domain-general follower moves (acknowledge, check, explain, query, reply) and a turn-boundary coefficient, plus move-specific and EOT activation directions estimated by shrinkage LDA on MapTask onset deltas and transferred zero-shot to FindTask and CReST. Realization uses norm-relative residual injection with realization-aware gating. Across three multimodal backbones, selection averages 0.60 accuracy; oracle realization averages 60.0% (about 10.4 points above the strongest baseline); end-to-end human-move match averages 46.9%, matching LoRA SFT (46.6%) and beating the unsteered backbone by 12.5 points. A complementary turn-boundary direction monotonically controls response length. Evaluation relies primarily on a prompted Qwen2.5-72B move classifier, validated against corpus gold and a 500-item three-annotator study.

Significance. Recovering an explicit, modular selection–realization interface inside frozen speech LLMs is a timely and practically relevant contribution: it reconnects modern end-to-end models to classical dialogue management without weight updates, and the MapTask-only direction transfer plus multi-backbone, multi-dataset design is stronger than typical steering papers. Matching SFT on move accuracy while remaining label-free at inference and modular is a concrete systems result. Strengths include dialogue-grouped splits, a broad baseline suite (prompt, PBL, SI, FUDGE, DeAL, PAS, SFT), ablations on modality/strength/layer/site, controllability matrices, and a blinded human study with method-ranking correlation (Spearman ρ=0.96). If the metric holds under stricter human scrutiny, the work offers a reusable control interface for grounding, repair, and floor management in spoken agents.

major comments (3)
  1. [Automatic Move Classifier; Tables 2, 4, 5] Automatic Move Classifier section and Tables 4–5: Oracle realization and end-to-end claims rest almost entirely on labels from one prompted Qwen2.5-72B few-shot classifier. The same classifier supplies the soft-lift objective used to choose per-move layers ℓ*_m and strengths α*_m on MapTask training contexts, then scores every method. Human validation (73.9% plurality agreement on 500 generated items; Fleiss' κ=0.682) is reported mainly as aggregate agreement and cross-method rank correlation (ρ=0.96), not as a per-method, per-move breakdown that would rule out differential bias toward short, formulaic surface forms that steering+EOT gating preferentially produce (acknowledge/query/reply). Please add (i) human majority accuracy by method (at least Base, strongest baseline, Latent-IM, SFT) and move type on the 500-item set or an expanded sample, and (ii) a short bias analysis (e.g., lengt
  2. [Method (Realization-aware gating); Table 11; Analysis] Realization-aware gating (Method; Analysis): Move steering is withheld for reply and EOT steering is applied only to acknowledge/query/reply, with gates driven by controller predictions and training-set steerability (reply controllability 0.14 vs query 0.99). This is well motivated by the controllability matrix (Figure 3) and improves E2E realization, but the policy is a free design choice fitted where steering lands. Report end-to-end Acc/BLEU with gating ablated (always-steer, never-steer, gold-move gates) on the same splits so readers can separate the contribution of the learned directions from the contribution of the hand-specified gate table (Table 11). If ungated reply steering collapses accuracy, state that limit explicitly in the main results rather than only in Analysis.
  3. [Abstract; Introduction; Results (End-to-End); Analysis; Conclusion] Task & Data / Future Work: The central objective is matching held-out human move labels. The Analysis section correctly notes irreducible follower choice under ambiguity (mode probabilities 0.30–0.51 after instructions; controller tracks human entropy at r=0.95) and that task-optimal moves may differ (MapTask map-mismatch example). That is fine for an imitation study, but Tables 4–5 and the abstract frame results as “appropriate” move control and dialogue management. Please tighten claim language in Abstract/Introduction/Conclusion to “human-move imitation / match” wherever task success is not measured, and briefly state that no task-outcome metric (route accuracy, object find rate, etc.) is reported—so gains should not be read as improved task success.
minor comments (5)
  1. [Table 5; End-to-End Move Control] Table 5: Latent-IM often trails SFT on BLEU while matching or beating move Acc. A short paragraph interpreting this (move-correct but lexically diverse realizations vs. reference imitation) would help readers who treat BLEU as primary quality.
  2. [Figure 1; Eqs. (5), (8); Steering Implementation] Figure 1 and Eq. (5)/(8): Clarify whether steering is applied only at decoding steps or also during prefill, and whether α/β are constant across tokens (Implementation Details say constant; main text is slightly less explicit).
  3. [Table 3] Table 3: Markov-1 receives the gold previous move; please mark this more visibly as an oracle-history upper reference so it is not read as a fair deployable baseline.
  4. [Throughout] Minor typos/spacing artifacts from PDF extraction appear throughout (e.g., “WeaskwhetheranLLM-internalanalogue”, “movefromthedialoguecontext”). Clean these in the camera-ready source.
  5. [Related Work] Related Work: A one-sentence contrast with full-duplex / interruption steering (Chang et al. 2026; Yegorova et al. 2026, already cited) on what is move-level vs. channel/turn-level control would sharpen positioning.

Circularity Check

1 steps flagged

No definitional circularity: selection, directions, and steering gains are empirical fits evaluated on held-out human moves, not inputs renamed as predictions.

specific steps
  1. fitted input called prediction [Steering Implementation; Layer and strength selection; Tables 4–5]
    "Stage 1 evaluates all layers at a fixed α=0.1 and selects the layer maximizing the mean paired soft lift Ec[Psteered(m|c)−Pbase(m|c)], where … P is obtained from the Qwen2.5-72B move classifier. Stage 2 fixes the selected layer and sweeps α∈{0.05,0.1,0.2,0.3,0.4}, choosing the value with the greatest soft lift"

    Per-move layer and α are fitted to maximize the same automatic move-classifier probability mass that later defines oracle-realization and end-to-end accuracy. This couples hyperparameter search to the reported metric and can inflate gains if the judge favors steered surface forms; it is not full by-construction circularity (held-out contexts and causal generation still intervene), so it warrants only a minor flag.

full rationale

Latent-IM’s chain is standard supervised/causal ML, not a first-principles derivation that collapses into its premises. Move selection trains a sidecar on gold yt via cross-entropy over residual-stream features and is scored against held-out gold (Table 3). Move directions are Fisher/LDA contrasts of gold-labeled onset deltas Δℓ_t on MapTask train, then transferred unchanged; oracle and end-to-end success are measured by whether generated text matches held-out human moves (via a separate few-shot classifier, with human plurality checks). That is not X-defined-as-Y: LDA on human onset activations does not by construction force generated tokens to receive the target label. Turn-boundary control is a simple yield−hold mean difference with a monotonic length sweep. Self-citations (Avsian/Dokme/Heck arXiv notes) are background, not uniqueness theorems that forbid alternatives. The only mild methodological coupling is that layer/α are chosen by classifier soft-lift on train, and headline realization/E2E numbers use the same judge family—an evaluation-validity concern, not a by-construction identity of prediction with fitted input. Score 1.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

The work is empirical systems ML. It assumes frozen multimodal LLMs already encode dialogue state and move intent in residual streams, that five mapped dialogue acts are the right control vocabulary, that human follower moves are the right supervision target, and that a prompted 72B classifier is adequate for evaluation and hyperparameter selection. Free parameters are the usual steering/controller knobs (layers, α, β calibration, PCA rank, gates). No new physical entities; invented constructs are methodological (move subspaces, turn-boundary direction, Latent-IM pipeline).

free parameters (5)
  • Per-move steering layer ℓ*_m and strength α*_m = α sweep {0.05..0.4} after layer pick at α=0.1; final defaults per move/backbone
    Chosen by two-stage sweep on MapTask training soft lift and degeneracy rate; transferred fixed to all eval sets.
  • Controller layer ℓ*_ctrl and SSM/readout weights = ~0.5M controller; layer selected for readout
    Trained with CE + smooth-L1 on gold moves/lengths; architecture and layer are design choices fitted to data.
  • Turn-boundary coefficient β and length calibration = β swept [-1,1]; mean length 71.3→10.4 words
    Monotone map from predicted log-length to β; gating applies only to terse moves.
  • PCA dimension k and Ledoit–Wolf LDA shrinkage for move directions = k≤128
    k=min(128,n-1,d); automatic shrinkage in one-vs-rest Fisher directions.
  • Realization-aware gating policy = Table 11 binary gates
    Hand/rule policy from measured steerability: withhold move steer for reply; withhold EOT for check/explain.
axioms (6)
  • domain assumption Frozen speech-LLM residual streams contain linearly readable dialogue state sufficient for next-move prediction without backbone updates.
    Foundation of the controller; supported empirically in Table 3 but not proved generally.
  • domain assumption Five mapped acts {acknowledge, check, explain, query, reply} are an adequate domain-general move inventory for follower turns in these tasks.
    Table 1 taxonomy and corpus label maps; excludes many corpus acts from targets.
  • ad hoc to paper Matching held-out human move labels is the right objective for 'appropriate' move control in this paper.
    Stated use of control 'to reproduce human move choices'; Future Work notes task utility may differ.
  • domain assumption Onset residual delta Δ=h(first response token)−h(previous) identifies a causal move subspace under additive steering.
    Eq. 3–5; standard steering assumption, validated by controllability matrix not by causal identifiability proof.
  • ad hoc to paper Qwen2.5-72B few-shot move classification is a valid primary metric for generated responses.
    Automatic Move Classifier section; partially validated by human study.
  • standard math Standard linear algebra / LDA / SSM sequence modeling results hold for these activation features.
    Fisher directions, PCA, selective SSM updates (Gu & Dao).
invented entities (3)
  • Latent-IM (selection + realization control interface) no independent evidence
    purpose: Name the end-to-end internal dialogue-management framework on frozen speech LLMs.
    Methodological system, not a physical entity; evaluated only inside this paper's experiments.
  • Move-specific residual-stream directions v_m (Fisher/LDA onset subspaces) no independent evidence
    purpose: Causally realize acknowledge/check/explain/query/reply via activation addition.
    Estimated per backbone on MapTask; transfer tested on other corpora but no external biological/linguistic grounding.
  • Turn-boundary direction v_eot no independent evidence
    purpose: Graded control of when the model yields the floor / response length.
    Difference of yield vs hold mean-pooled activations; predictive AUROC reported but entity is internal construct.

pith-pipeline@v1.2.0-daily-grok45 · 24383 in / 4056 out tokens · 80252 ms · 2026-07-30T16:43:52.461349+00:00 · methodology

0 comments
read the original abstract

Classical spoken dialogue systems often separated dialogue management from response realization: a policy selected the next dialogue action, and a generation component expressed that action. As dialogue systems shift toward LLMs, this decomposition has largely disappeared into the model's hidden representations. We ask whether an LLM-internal analogue of state estimation and action control can be recovered for conversational moves such as acknowledging, checking, querying, explaining, and replying. We formulate move control as two coupled problems: selection, predicting the appropriate next move from the dialogue context, and realization, causally producing a chosen move at generation time. We introduce Latent-IM, an internal dialogue-management framework that provides a general interface for choosing and deploying conversational moves under different objectives. Here, we use this control to reproduce human move choices, improving average end-to-end move accuracy by 12.5 points over the unsteered backbone while performing comparably to fine-tuning.

Figures

Figures reproduced from arXiv: 2607.26928 by Adar Avsian, Atahan Dokme, Larry Heck, Tony Woo.

Figure 1
Figure 1. Figure 1: (a) Before generation, the controller predicts the next conversational move, whether and how strongly to apply move [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Steering along the turn-boundary direction mono [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Controllability matrix for move steering (map-task, [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The controller reads the frozen backbone’s residual [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Human annotation interface for move classification. Annotators are shown the preceding dialogue context and a single [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 13 linked inside Pith

  1. [1]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    Grounding gaps in language model generations , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2024

  2. [2]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    Reflect, not reflex: Inference-based common ground improves dialogue response quality , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    2022

  3. [3]

    Fourth Workshop on Bridging Human--Computer Interaction and Natural Language Processing , pages=

    Dialogue Acts as a Lens on Human--LLM Interaction: Analyzing Conversational Norms in Model-Generated Responses , author=. Fourth Workshop on Bridging Human--Computer Interaction and Natural Language Processing , pages=

  4. [4]

    arXiv preprint arXiv:2512.21120 , year=

    ClarifyMT-Bench: Benchmarking and Improving Multi-Turn Clarification for Conversational Large Language Models , author=. arXiv preprint arXiv:2512.21120 , year=

    arXiv

  5. [5]

    Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue , pages=

    Controllable generation of dialogue acts for dialogue systems via few-shot response generation and ranking , author=. Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue , pages=

  6. [6]

    Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue , pages=

    DiactTOD: Learning generalizable latent dialogue acts for controllable task-oriented dialogue systems , author=. Proceedings of the 24th Annual Meeting of the Special Interest Group on Discourse and Dialogue , pages=

  7. [7]

    ACM Transactions on Information Systems , volume=

    Proactive conversational ai: A comprehensive survey of advancements and opportunities , author=. ACM Transactions on Information Systems , volume=. 2025 , publisher=

    2025

  8. [8]

    , author=

    Grounding in communication. , author=. 1991 , publisher=

    1991

  9. [9]

    Computational intelligence , volume=

    Conversation acts in task-oriented spoken dialogue , author=. Computational intelligence , volume=. 1992 , publisher=

    1992

  10. [10]

    URL https://arxiv

    Steering language models with activation engineering, 2024 , author=. URL https://arxiv. org/abs/2308.10248 , volume=

    Pith/arXiv arXiv 2024

  11. [11]

    arXiv preprint arXiv:2410.16314 , year=

    Steering large language models using conceptors: Improving addition-based activation engineering , author=. arXiv preprint arXiv:2410.16314 , year=

    Pith/arXiv arXiv

  12. [12]

    language , volume=

    A simplest systematics for the organization of turn-taking for conversation , author=. language , volume=. 1974 , publisher=

    1974

  13. [13]

    Language and speech , volume=

    The HCRC map task corpus , author=. Language and speech , volume=. 1991 , publisher=

    1991

  14. [14]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

  15. [15]

    Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations , pages=

    Transformers: State-of-the-art natural language processing , author=. Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations , pages=

    2020

  16. [16]

    and Sengupta, Sailik and Bonadiman, Daniele and Lai, Yi-An and Gupta, Arshit and Pappas, Nikolaos and Mansour, Saab and Kirchhoff, Katrin and Roth, Dan , year=

    Huang, James Y. and Sengupta, Sailik and Bonadiman, Daniele and Lai, Yi-An and Gupta, Arshit and Pappas, Nikolaos and Mansour, Saab and Kirchhoff, Katrin and Roth, Dan , year=. DeAL: Decoding-time Alignment for Large Language Models , url=. doi:10.18653/v1/2025.acl-long.1274 , booktitle=

    doi:10.18653/v1/2025.acl-long.1274 2025

  17. [17]

    2026 , eprint=

    Painless Activation Steering: An Automated, Lightweight Approach for Post-Training Large Language Models , author=. 2026 , eprint=

    2026

  18. [18]

    2023 , eprint=

    Controllable Generation of Dialogue Acts for Dialogue Systems via Few-Shot Response Generation and Ranking , author=. 2023 , eprint=

    2023

  19. [19]

    2025 , eprint=

    Improving Instruction-Following in Language Models through Activation Steering , author=. 2025 , eprint=

    2025

  20. [20]

    FUDGE: Controlled Text Generation With Future Discriminators , url=

    Yang, Kevin and Klein, Dan , year=. FUDGE: Controlled Text Generation With Future Discriminators , url=. doi:10.18653/v1/2021.naacl-main.276 , booktitle=

    doi:10.18653/v1/2021.naacl-main.276 2021

  21. [21]

    Multimodality and Dialogue Act Classification in the R obo H elper Project

    Chen, Lin and Di Eugenio, Barbara. Multimodality and Dialogue Act Classification in the R obo H elper Project. Proceedings of the SIGDIAL 2013 Conference. 2013

    2013

  22. [22]

    The I ndiana ``Cooperative Remote Search Task'' ( CR e ST ) Corpus

    Eberhard, Kathleen and Nicholson, Hannele and K. The I ndiana ``Cooperative Remote Search Task'' ( CR e ST ) Corpus. Proceedings of the Seventh International Conference on Language Resources and Evaluation ( LREC '10). 2010

    2010

  23. [23]

    arXiv preprint arXiv:2603.29846 , year=

    SNEAK: Evaluating Strategic Communication and Information Leakage in Large Language Models , author=. arXiv preprint arXiv:2603.29846 , year=

    arXiv

  24. [24]

    Advances in Neural Information Processing Systems , volume=

    Inference-time intervention: Eliciting truthful answers from a language model , author=. Advances in Neural Information Processing Systems , volume=

  25. [25]

    International conference on learning representations , volume=

    Function vectors in large language models , author=. International conference on learning representations , volume=

  26. [26]

    arXiv preprint arXiv:2312.06681 , year=

    Steering llama 2 via contrastive activation addition , author=. arXiv preprint arXiv:2312.06681 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2601.04157 , year=

    FLEx: Language Modeling with Few-shot Language Explanations , author=. arXiv preprint arXiv:2601.04157 , year=

    arXiv

  28. [28]

    Computer Speech & Language , volume=

    Partially observable Markov decision processes for spoken dialog systems , author=. Computer Speech & Language , volume=. 2007 , publisher=

    2007

  29. [29]

    Proceedings of the IEEE , volume=

    Pomdp-based statistical spoken dialog systems: A review , author=. Proceedings of the IEEE , volume=. 2013 , publisher=

    2013

  30. [30]

    arXiv preprint arXiv:2312.00752 , year=

    Mamba: Linear-time sequence modeling with selective state spaces , author=. arXiv preprint arXiv:2312.00752 , year=

    Pith/arXiv arXiv

  31. [31]

    Annals of eugenics , volume=

    The use of multiple measurements in taxonomic problems , author=. Annals of eugenics , volume=. 1936 , publisher=

    1936

  32. [32]

    Journal of multivariate analysis , volume=

    A well-conditioned estimator for large-dimensional covariance matrices , author=. Journal of multivariate analysis , volume=. 2004 , publisher=

    2004

  33. [33]

    2025 , eprint=

    Qwen2.5-Omni Technical Report , author=. 2025 , eprint=

    2025

  34. [34]

    arXiv preprint arXiv:2509.17765 , year=

    Qwen3-omni technical report , author=. arXiv preprint arXiv:2509.17765 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2503.01743 , year=

    Phi-4-mini technical report: Compact yet powerful multimodal language models via mixture-of-loras , author=. arXiv preprint arXiv:2503.01743 , year=

    Pith/arXiv arXiv

  36. [36]

    Findings of the Association for Computational Linguistics: ACL 2022 , pages=

    Extracting latent steering vectors from pretrained language models , author=. Findings of the Association for Computational Linguistics: ACL 2022 , pages=

    2022

  37. [37]

    arXiv preprint arXiv:2310.01405 , year=

    Representation engineering: A top-down approach to ai transparency , author=. arXiv preprint arXiv:2310.01405 , year=

    Pith/arXiv arXiv

  38. [38]

    Language , volume=

    The preference for self-correction in the organization of repair in conversation , author=. Language , volume=. 1977 , publisher=

    1977

  39. [39]

    Proceedings of the Second SIGdial Workshop on Discourse and Dialogue , year=

    On the means for clarification in dialogue , author=. Proceedings of the Second SIGdial Workshop on Discourse and Dialogue , year=

  40. [40]

    Transactions of the Association for Computational Linguistics , volume=

    Generative spoken dialogue language modeling , author=. Transactions of the Association for Computational Linguistics , volume=. 2023 , publisher=

    2023

  41. [41]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Beyond turn-based interfaces: Synchronous llms as full-duplex dialogue agents , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  42. [42]

    Behavior-SD: Behaviorally aware spoken dialogue generation with large language models , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  43. [43]

    Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue , pages=

    Towards a general, continuous model of turn-taking in spoken dialogue using LSTM recurrent neural networks , author=. Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue , pages=

  44. [44]

    Findings of the Association for Computational Linguistics: EMNLP 2020 , pages=

    TurnGPT: a transformer-based language model for predicting turn-taking in spoken dialog , author=. Findings of the Association for Computational Linguistics: EMNLP 2020 , pages=

    2020

  45. [45]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Predicting Turn-Taking and Backchannel in Human-Machine Conversations Using Linguistic, Acoustic, and Visual Signals , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  46. [46]

    Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue , pages=

    Attentive listening system with backchanneling, response generation and flexible turn-taking , author=. Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue , pages=

  47. [47]

    arXiv preprint arXiv:2606.11386 , year=

    Overcoming State Inertia in Full-Duplex Spoken Language Models via Activation Steering , author=. arXiv preprint arXiv:2606.11386 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2606.11400 , year=

    Steering Where to Listen: Instruction-Based Activation Steering Redirects Temporal Attention in Large Audio-Language Models , author=. arXiv preprint arXiv:2606.11400 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2606.00460 , year=

    SALSA: Speech Aware LLM Adaptation via Learned Steering Activation Vectors , author=. arXiv preprint arXiv:2606.00460 , year=

    Pith/arXiv arXiv

  50. [50]

    Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

    Bleu: a method for automatic evaluation of machine translation , author=. Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

  51. [51]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  52. [52]

    Advances in neural information processing systems , volume=

    Training language models to follow instructions with human feedback , author=. Advances in neural information processing systems , volume=

  53. [53]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  54. [54]

    2026 , url =

    Dokme, Atahan and Reichman, Benjamin and Heck, Larry , booktitle =. 2026 , url =

    2026

  55. [55]

    arXiv preprint arXiv:2607.19326 , year =

    Selective State-Space Adaptation and Retrieval for Language Model Reasoning , author =. arXiv preprint arXiv:2607.19326 , year =

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2511.05963 , year=

    Next-latent prediction transformers learn compact world models , author=. arXiv preprint arXiv:2511.05963 , year=

    Pith/arXiv arXiv