icassp2027-experiment/manuscript/main/main.tex

% ICASSP 2027 content skeleton. The final paper must use the official 2027
% author kit and contain no visible \todo{} markers. Planned experiments are
% intentionally marked rather than presented as completed work.
\documentclass{article}
\usepackage{spconf}
\usepackage{amsmath,amssymb,graphicx,hyperref,xcolor}

\newcommand{\todo}[1]{\textcolor{red}{\textbf{[TODO: #1]}}}
\newcommand{\vect}[1]{\boldsymbol{#1}}
\setlength{\emergencystretch}{2em}

\title{Is It My Turn? Causal Local Evidence Adaptation of Omni Models for
Egocentric Floor Forecasting}

% TODO: Insert the final author list and affiliations after the ICASSP 2027
% review policy and author order are finalized.
\name{Internal Manuscript Draft}
\address{}

\begin{document}
\ninept
\maketitle

\begin{abstract}
Real-time audiovisual agents must decide not only what to say, but whether an
emerging floor opportunity belongs to them. This is difficult in multiparty
egocentric interaction: the target wearer is behind the camera, while the
visible people provide only indirect evidence about whether the surrounding
floor remains occupied. We study strictly causal forecasting of the wearer's
future speech activity every 200~ms. Although a general-purpose Omni model
provides rich long-context audio--visual representations, its global readout is
not explicitly trained to expose sub-second, wearer-relative floor evidence.
We therefore study a candidate causal local floor adapter that combines the global
Omni state with short-window mixed audio and an anonymous set of visible-person
observations, without assuming persistent speaker identity. On EasyCom,
\todo{report the final three-seed dense and fixed-horizon behavioral-event
results: wearer-entry and INT-yield AP, recall-matched false-positive rates,
confidence intervals, and oracle-gap recovery}. Controlled
\todo{time-shift/remove-vision/same-rate or capacity-matched} interventions
show that \todo{state the supported mechanism, without claiming addressee or
speaker identity}. These results demonstrate \todo{the narrow conclusion
actually supported by the final experiments}.
\end{abstract}

\begin{keywords}
conversational agents, turn-taking, egocentric video, multimodal adaptation,
voice activity projection
\end{keywords}

\section{Introduction}
\label{sec:intro}

Natural turn-taking requires a conversational system to anticipate not only
when speech will stop, but whether a floor opportunity actually belongs to it.
In a dyadic exchange, a pause after one interlocutor finishes often invites the
other to respond. In a group conversation, the same pause may instead mark a
handover between two other participants. An egocentric agent that reacts only
to silence or aggregate speech may therefore interrupt a conversation that
should continue without it. We study \emph{egocentric floor forecasting}:
continuously predicting the near-future speech participation of the
first-person wearer from strictly causal audio--visual observations.

Voice Activity Projection (VAP) introduced continuous prediction of future
voice activity for turn-taking \cite{ekstedt2022vap}. EgoSpeak brought this
formulation to causal, untrimmed first-person streams, using future camera-
wearer activity as an imitation target \cite{kim2025egospeak}. Complementary
work adds visual gaze, head pose, and facial behavior to predictive turn-taking
\cite{russell2025mmvap}, while MuVAP combines global acoustic dynamics with
visible face tracks to forecast visible floor holders in multiparty scenes
\cite{qi2026muvap}. Our target has a different role structure: the participant
whose future action is predicted is the one participant absent from the
visible set. The visible people are contextual evidence for the unseen
wearer's decision, not candidate output identities.

This role asymmetry exposes an adaptation problem for general-purpose Omni
models. Omni pretraining favors globally integrated representations useful for
multimodal understanding and generation. Egocentric floor control instead
depends on timely prosodic and visible-person activity at sub-second scale,
including whether other participants continue to occupy the floor. Increasing
model capacity or applying a generic parameter-efficient update does not
explicitly preserve or route this evidence to a dense wearer-relative
prediction head. The open question is therefore not merely whether an Omni
backbone can be applied to EgoSpeak, but what task structure is still required
to turn general audiovisual understanding into reliable floor control.

Our candidate decomposition separates long-context understanding from
short-window floor evidence. A
causal global Omni state represents the broader scene and conversation. In
parallel, a lightweight local path encodes recent mixed audio and a changing,
anonymous set of visible-person observations. A residual adapter makes this
local evidence accessible to the common ten-horizon activity head. The primary
variant deliberately avoids cross-time identity: the current development
oracle supports timely aggregate visible-other activity but not person
continuity. Track-aware structure is considered only if it defeats no-ID,
identity-corrupted, and capacity-matched controls.

The final paper will claim the following three contributions only if the
corresponding TODO gates pass:
\begin{itemize}
  \item We identify and test a temporal--role mismatch between a global Omni
  readout and causal egocentric floor forecasting for an invisible wearer.
  \item We evaluate a wearer-relative causal local floor adapter that combines
  long-context Omni information with short-window audio and anonymous
  visible-person evidence.
  \item We evaluate whether the adapter assigns lower erroneous-action scores
  during multiparty floor transfers under an entry/INT protocol with point
  positives and explicitly annotated hard-negative opportunity regions.
\end{itemize}

\begin{figure*}[t]
\centering
\includegraphics[width=0.98\textwidth]{figures/method_architecture.pdf}
\caption{Our strictly causal global--local architecture (top) and streaming
alignment (bottom). At every 5-Hz cutoff $t_n$, the global state includes the
latest partial 2-s Omni block, while every fifth causal 25-Hz local state
provides shared audio and an anonymous visible-person set. Masked person
attention contributes only through a zero-initialized residual to the matched
$G+U$ baseline.}
\label{fig:method}
\end{figure*}

\section{Method}
\label{sec:method}

\subsection{Model Architecture}
\label{sec:model_arch}

At decision time $t$, the model observes only the audiovisual stream
$\mathcal X_{\leq t}$ and predicts $K=10$ future bins spaced by
$\Delta=0.2$~s. For horizon $k$, the multilabel target
$\vect y_{t,k}=[y^{\mathrm{bg}},y^{\mathrm{w}},y^{\mathrm{o}}]$ denotes
background, camera-wearer activity, and activity by one or more other people;
wearer--other overlap is allowed. The output is
\begin{equation}
 \hat{\vect Y}_t=f_\theta(\mathcal X_{\leq t})
 \in[0,1]^{K\times3}.
 \label{eq:task}
\end{equation}

Our architecture complements globally integrated Omni context with explicitly
accessible local floor evidence (Fig.~\ref{fig:method}). Let
$\vect g_t\in\mathbb R^{d_g}$ denote the causal Qwen2.5-Omni Thinker readout
at $t$ \cite{xu2025qwenomni}. In parallel, the frozen MuVAP ASD frontend
\cite{qi2026muvap} produces one shared mixed-audio CPC state $\vect u_t$ and a
variable-size set of visible-person states
\begin{equation}
 \mathcal P_t=\{\vect p_{t,i}\}_{i=1}^{N_t},\qquad
 \vect p_{t,i}=[\vect v_{t,i};s_{t,i}],
 \label{eq:person_state}
\end{equation}
where $\vect v_{t,i}$ is a causal visual state and $s_{t,i}$ is one
audio--visual speaking logit. The shared audio state is not duplicated in
each person token. The prediction model receives neither participant IDs nor
cross-time track identity; provided causal face boxes are used only to isolate
the representation and fusion question.

A matched global--audio head first maps $[\vect g_t;\vect u_t]$ to baseline
logits $\vect Z_t^{\mathrm{base}}\in\mathbb R^{K\times3}$ and an intermediate
state $\vect h_t$. The same inputs query the current person set:
\begin{equation}
 \begin{aligned}
  \vect q_t &= W_g\vect g_t+W_a\vect u_t,\\
  \mathcal M_t &= \{W_p\vect p_{t,i}\}_{i=1}^{N_t}
                  \cup\{\vect p_\varnothing\},\\
  \vect c_t &= \mathrm{MHA}(\vect q_t,\mathcal M_t,\mathcal M_t),
 \end{aligned}
 \label{eq:person_attention}
\end{equation}
where masking supports variable cardinality and the learned null token
$\vect p_\varnothing$ makes the attention well defined for an empty set. If
$m_t$ indicates that at least one valid person observation exists, the person
branch supplies only a residual logit correction,
\begin{equation}
 \vect Z_t=\vect Z_t^{\mathrm{base}}
 {}+m_t\,\mathrm{reshape}\!\left(W_r[\vect h_t;\vect c_t;m_t]\right),
 \qquad W_r\vert_{\mathrm{init}}=0,
 \label{eq:residual}
\end{equation}
and $\hat{\vect Y}_t=\mathrm{softmax}_c(\vect Z_t)$. Thus, after copying and
freezing the trained $G+U$ head, the complete model initially reproduces that
baseline exactly; only the person projection, query, masked attention, and
residual readout are trained in the adaptation stage.

We use the EgoSpeak/LSTR multi-positive cross-entropy over causal-valid rows,
\begin{equation}
 \begin{aligned}
  \mathcal L&=-\frac{1}{M}\sum_{t,k,c}m_{t,k}y^c_{t,k}
  \log\mathrm{softmax}_c(\vect Z_{t,k}),\\
  M&=\sum_{t,k}m_{t,k}.
 \end{aligned}
 \label{eq:loss}
\end{equation}
For an overlap target, both active channels contribute a log-probability term.
No transcript, target-of-speech annotation, close-microphone signal, speech
label, or future observation is supplied to the forecasting model. The
wearer forecast is interpreted only at evaluation time: it is an entry proxy
when the wearer is silent and its complement is a yield proxy when the wearer
is active, not a trained event hazard or committed controller action.

\subsection{Causal Streaming Architecture}
\label{sec:causal_streaming}

We distinguish encoder processing granularity from the common forecasting
rate. Predictions are emitted at $t_n=n\Delta$, $n\geq1$, while every
representation used at $t_n$ is restricted to media available by that physical
cutoff.

The global path uses an 8-s rolling causal presentation budget with
source-aligned 2-s audiovisual blocks. Completed blocks are cached and
replayed, while the currently incomplete block is a temporary tail. At every
200-ms cutoff, the tail is rebuilt with all newly observed audio and video and
included in computing $\vect g_{t_n}$; when it reaches 2~s, it is committed to
the reusable history. Consequently, the Omni state is not held constant for
2~s: it includes the latest available content at every decision cutoff without
observing media after $t_n$. We retain the final four Thinker query-layer
states and fuse them in the matched head with a concatenation MLP.

The local frontend operates causally at 25~Hz. State $j$ consumes the
half-open interval $[j/25,(j+1)/25)$ and is timestamped at its right edge.
Mixed audio is encoded once as a continuous shared CPC stream, while each
available face crop is processed by the causal visual and audio--visual
speaking modules. Ten-second extraction chunks may prepend up to 2~s of past
context, never future context. At a 5-Hz forecasting cutoff, we select the
local state whose right edge equals $t_n$; hence every fifth 25-Hz state aligns
with one Omni prediction row. The two paths therefore describe the same
causal prefix: the Omni path preserves integrated conversational context,
whereas the local path exposes fine-grained audio and visible-person floor
evidence. Fusion and prediction occur only on the common 5-Hz grid.


\section{Experiments}
\label{sec:experiments}

\subsection{Dataset and Data Preparation}

We evaluate on EasyCom following the EgoSpeak split, using Sessions 4--12 as
the training/development pool and Sessions 1--3 only for final evaluation
\cite{donley2021easycom}. The participant-resolved speech annotations are
converted into two parallel representations. First, they are downsampled to
5~Hz to construct the EgoSpeak-compatible ten-horizon targets for background,
camera-wearer, and aggregate non-wearer activity. Second, the original
participant timeline is retained to derive adjudicated interaction events for
behavioral evaluation.

\textbf{Behavioral event annotation.}
We use every observed wearer activity transition as a high-recall semantic
candidate. For each wearer onset, nearby preceding non-wearer releases, the
intervening silence gap, and concurrent activity are retained as contextual
facts; the onset is labeled as wearer entry only when the subsequent wearer
speech constitutes a genuine floor take, including re-entry while another
speaker is active. For each wearer offset, nearby non-wearer onsets and offsets,
pre-existing and newly active speakers, overlap durations, and subsequent
speakers are retained without a timing-based takeover gate; the offset is
labeled as INT-yield only when a non-wearer genuinely takes over and the wearer
relinquishes the floor. We let an LLM jointly inspect
participant-resolved timing and transcripts to provide a semantic pre-label,
followed by mandatory human adjudication. Separately, we mine regions with no
wearer response that nevertheless constitute false-positive opportunities:
same-speaker pauses and other-to-other transitions for entry, and backchannels or
non-takeover overlaps for INT. Only human-confirmed positives and hard negatives
enter evaluation. Appendix~\ref{app:event_annotation} gives the full protocol.

\begin{table}[t]
\caption{Behavioral-event annotation statistics after human adjudication.
Counts must be filled only after the candidate protocol and annotation set are
frozen. Scoring-eligible totals include confirmed positives and hard negatives
only.}
\label{tab:behavior_annotation_stats}
\centering
\small
\begin{tabular}{lcc}
\hline
Annotation outcome & Train/dev & Test \\
\hline
\multicolumn{3}{l}{\emph{Wearer entry}}\\
Positive: wearer entry              & \todo{--} & \todo{--} \\
HN: same-speaker pause              & \todo{--} & \todo{--} \\
HN: other-to-other transition       & \todo{--} & \todo{--} \\
HN: wearer backchannel              & \todo{--} & \todo{--} \\
\hline
\multicolumn{3}{l}{\emph{INT-yield}}\\
Positive: INT-yield                 & \todo{--} & \todo{--} \\
HN: other backchannel               & \todo{--} & \todo{--} \\
HN: non-takeover overlap            & \todo{--} & \todo{--} \\
HN: wearer holds floor              & \todo{--} & \todo{--} \\
\hline
Excluded: valid but out of scope    & \todo{--} & \todo{--} \\
Excluded: insufficient evidence     & \todo{--} & \todo{--} \\
Excluded: data quality              & \todo{--} & \todo{--} \\
\textbf{Scoring-eligible total}     & \todo{--} & \todo{--} \\
\hline
\end{tabular}
\end{table}

Participant identity is used only to construct these evaluation annotations
and is never provided to the prediction model. Provided person boxes are used
only for the causal visible-person input; transcripts, target-of-speech labels,
close-microphone audio, and future observations are unavailable to the model.
\todo{Report duration and event statistics: entry positives/hard negatives,
INT positives/hard negatives, excluded candidates, and audited-region counts
for train/development and test. Freeze the event-construction and region-audit
protocol before reporting semantic entry/INT results; until then, retain the
timing-derived proxy terminology.}

\subsection{Model Configuration and Training}

Our main backbone is Qwen2.5-Omni-3B. We compare a frozen global Omni readout,
matched PEFT/LoRA adaptation, global features with recent local audio, a
parameter- and history-matched temporal local branch, and the proposed causal
local audio--visual adapter. Qwen2.5-Omni-7B is used as a capacity-scaling
control. All learned variants use identical causal examples, prediction
horizons, prediction heads, and the multi-positive cross-entropy objective in
Eq.~\eqref{eq:loss}.

The diagnostic streaming ablation additionally uses MiniCPM-o 4.5
\cite{cui2026minicpmo45}. Its native cached stream is evaluated at its native
one-second update interval: a state becomes available to the 200-ms row whose
cutoff equals the chunk end and is then causally held until the next update. A
Qwen row subjected to the same 1-Hz update-and-hold schedule separates
refresh-rate effects from backbone effects. A MiniCPM-o
5-Hz cache row is admissible only if a cached branch at every cutoff agrees
with full causal-prefix recomputation within a frozen numerical tolerance and
passes future-input intervention tests. Thus the native 1-Hz condition is a
strictly causal held-state baseline, not a claim of native 5-Hz refresh.

Model selection is performed using training-side data only. Event operating
points are selected from session-level out-of-fold predictions over
Sessions~4--12; the final model is then retrained on all nine sessions before
evaluation on Sessions~1--3. Final results are averaged over three random
seeds.
\todo{Freeze context length, local window/rates, optimizer, epochs, LoRA
configuration, trainable parameter counts, and inference latency.}

\subsection{Evaluation Protocol}

We first retain dense future-activity evaluation for direct comparison with
EgoSpeak, reporting AP at the 0.2--2.0~s horizons and their mean mAP.

For dense-only models, behavioral evaluation begins with an explicitly
activity-derived event proxy rather than treating wearer voice activity as a
semantic event detector. Following the prediction-region principle used by
VAP~\cite{ekstedt2022vap}, every adjudicated item contributes one score from a
fixed 0.6-s (three-frame) causal decision region. A positive uses the three
source frames immediately preceding its corrected behavior anchor. An
independently mined hard negative uses the first three source frames after its
opportunity begins; a raw wearer boundary relabeled as a hard negative instead
uses the same pre-boundary region as a positive. Items without a complete
three-frame region are excluded and counted. The longer interval supplied to
the annotator is evidence context only and is never a model-scoring window.

At each decision frame, we follow the target-aligned forecast tensor
diagonally to recover the wearer probabilities emitted at that same causal
source time. We average the 0.6--2.0-s horizons to obtain a sustained wearer-
activity score. This mean is the entry proxy, and one minus the mean is the
INT-yield proxy. Max-pooling the three decision-frame scores gives exactly one
score per positive event or hard-negative opportunity. This construction can
capture sustained activity versus a brief vocalization, but it cannot by
itself establish semantic floor intent; in particular, wearer backchannels and
within-turn pauses remain deliberately difficult negatives. A later linear
event probe, if used, is fit only on training-side annotations and compared at
matched capacity across representations, following the separated downstream-
probe logic of MuVAP~\cite{qi2026muvap}.

\textbf{False-positive scoring rule.}
A false positive is a hard-negative opportunity whose fixed-region proxy score
exceeds the frozen threshold, regardless of whether the wearer subsequently
exhibits an onset or offset. Max-pooling means that multiple high-scoring
decision frames still count as one false-positive opportunity. Thus, false-positive
evaluation includes conversational situations in which the correct wearer
behavior is to remain silent or retain the floor.

We report threshold-free proxy AP for entry and INT. The headline behavioral
operating point additionally reports false-positive rate at matched recall
$R=0.8$; thresholds are selected independently for entry and INT from training-
side out-of-fold predictions. The other-to-other hard-negative subset tests
whether local visual evidence specifically reduces erroneous wearer action
scores during multiparty floor transfers. \todo{Freeze the 0.6-s decision
window and 0.6--2.0-s aggregation range using training-side data, run window
sensitivity analysis, and verify identical event denominators across methods.}

A deployment-style committed-event analysis with one-to-one matching and
signed latency is secondary and is omitted unless a threshold, persistence,
refractory period, matching interval, and competing-risk rule are all frozen
before test evaluation. The fixed-horizon protocol itself makes no detection-
latency claim.


\section{Results and Analysis}
\label{sec:results}

\subsection{Main Comparison}

Table~\ref{tab:main} compares the proposed causal local audiovisual adapter
with an EgoSpeak reproduction, the frozen Qwen2.5-Omni-3B global state $G$,
matched PEFT, $G$ plus recent local audio, and a parameter- and history-matched
temporal local branch. Qwen2.5-Omni-7B is included only as a capacity-scaling
control. All methods use the same split, causal inputs, targets, and evaluation
protocol. Event thresholds are selected on training-side calibration data and
frozen before test evaluation. \todo{Freeze all baseline implementations and
verify identical targets, causal validity, and evaluation.}

\begin{table}[t]
\caption{Confirmatory EasyCom results. Entry and INT false-positive rates use
the fixed-region activity-derived proxy at matched recall $R=0.8$; E/I AP is
the corresponding event-balanced proxy AP.}
\label{tab:main}
\centering
\small
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lccccc}
\hline
Method & 0.2-s mAP $\uparrow$ & Mean mAP $\uparrow$
& Entry FPR@$R=.8$ $\downarrow$ & INT FPR@$R=.8$ $\downarrow$ & E/I AP $\uparrow$\\
\hline
EgoSpeak reproduction       & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
Omni-3B frozen $G$          & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
Omni-3B matched PEFT        & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
$G+$local audio             & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
$G+$matched temporal local  & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
$G+$causal local AV (ours)  & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
Omni-7B scaling control     & \todo{--} & \todo{--} & \todo{--} & \todo{--} & \todo{--/--}\\
\hline
\end{tabular}}
\end{table}

\textbf{Dense forecasting.}
The proposed local AV adapter improves 0.2-s mAP from
\todo{$X$ to $Y$} over the strongest matched baseline and improves
mean mAP by \todo{$Z$ points}.
The gain is largest at \todo{short prediction horizons / transition-adjacent
frames / specify actual observation}, indicating that local audiovisual
evidence primarily improves \todo{the interpretation supported by the data}.

\textbf{Annotated behavior.}
At matched wearer-entry recall $R=0.8$, the proposed adapter reduces entry FPR
from \todo{$X$} to \todo{$Y$}; at matched INT-yield recall, it changes INT FPR
from \todo{$A$} to \todo{$B$}. Its entry/INT activity-proxy APs are
\todo{$E/I$}. These recall-matched opportunity error rates test whether lower
scores on confusing floor situations are obtained without suppressing
legitimate entry or yield events.

\textbf{Comparison with adaptation baselines.}
Matched PEFT yields \todo{result}, whereas adding recent local audio alone
yields \todo{result}.
The parameter- and history-matched temporal branch achieves
\todo{result}.
The remaining gap between this control and the proposed adapter therefore
suggests that \todo{correctly timed audiovisual local evidence contributes
beyond additional temporal capacity}.
\todo{If this difference is not statistically supported, state instead that
the local temporal branch accounts for most of the gain.}

\subsection{Information Opportunity}

Before designing the deployable adapter, we measured whether local floor
information was useful at all using an analysis-only oracle.
The oracle receives perfectly timed visible-other speech activity through a
non-deployable pathway and is never used in the main comparison.

On development Session~12, aggregate visible activity improved 0.2-s mAP by
6.97 points and ten-horizon mean mAP by 2.91 points over the matched
global-plus-local-audio baseline.
In contrast, visibility alone and one-second-delayed activity remained close
to the audio baseline, while person-specific histories did not outperform
identity-shuffled histories.

These controls indicate that the useful signal is primarily
\emph{timely local floor evidence}, rather than participant identity or
static visibility.
The proposed adapter is therefore designed to approximate this information
using only causally available audiovisual observations.

\todo{After completing all seeds, report mean $\pm$ confidence interval here.
If the oracle effect is unstable, describe it qualitatively and omit the exact
numbers from the four-page version.}

\subsection{Mechanism Analysis}

To determine whether the gain arises from correctly aligned visual floor
evidence rather than additional parameters or generic visual context, and to
separate streaming refresh from representation family, we evaluate the
controls in Table~\ref{tab:ablation}.

\begin{table}[t]
\caption{Mechanism and streaming/cache controls. The streaming panel uses
frozen global states and capacity-matched downstream heads; all rows share
targets and evaluation. FPRs are measured at matched recall $R=0.8$.
Cross-family raw scores are diagnostic.}
\label{tab:ablation}
\centering
\small
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccc}
\hline
Condition & Mean mAP $\uparrow$ & Entry FPR $\downarrow$
& O$\rightarrow$O FPR $\downarrow$ & INT FPR $\downarrow$\\
\hline
\multicolumn{5}{l}{\emph{Local-evidence controls (Qwen2.5-Omni-3B)}}\\
Remove vision           & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
Time-shift vision       & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
Full-frame vision       & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
Correct local AV        & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
\hline
\multicolumn{5}{l}{\emph{Streaming/cache controls}}\\
Qwen, full causal prefix at 5 Hz (no cache)
                         & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
Qwen, cached 5-Hz update & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
Qwen, cached 1-Hz + causal hold & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
MiniCPM-o, native cached 1-Hz + causal hold
                         & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
MiniCPM-o, audited cached 5-Hz branch
                         & \todo{--} & \todo{--} & \todo{--} & \todo{--}\\
\hline
\end{tabular}}
\end{table}

Removing local vision changes \todo{entry/O$\rightarrow$O/INT FPR by $X/Y/Z$},
while shifting the visual stream by \todo{$\Delta t$} produces \todo{result}.
Replacing person-local evidence with full-frame vision produces \todo{result}.
If aligned local AV improves estimates of surrounding floor occupancy, its
largest selective reduction should appear on the O$\rightarrow$O hard-negative
subset. Together, these comparisons \todo{support / do not support} that
mechanism beyond generic visual context.

\todo{Report paired intervals and the supported interpretation. If correct
and shifted/removed vision do not separate, drop the audiovisual
evidence-routing claim. If dense mAP improves but event behavior does not, do
not claim improved floor control. If legitimate onset or yield degrades, do
not describe lower false entry as success.}

\subsection{Scaling and Streaming Controls}

Increasing the frozen Omni backbone from 3B to 7B yields
\todo{$X$ points} in 0.2-s mAP and \todo{entry/INT FPR of $Y/Z$} at matched
recall.
The proposed 3B adapter \todo{remains better than / approaches / is surpassed
by} the 7B scaling control.

This result suggests that \todo{architectural access to local causal evidence
provides a complementary benefit to backbone scaling}.
Because both backbones belong to the Qwen2.5-Omni family, we do not claim that
this conclusion generalizes to all omni-modal models.

The streaming rows answer two different questions. First, full causal-prefix
recomputation and streaming cache reuse at the same 5-Hz Qwen cutoffs test
whether cache/chunk semantics change the representation; equivalent outputs
make this an efficiency control rather than a distinct predictive method.
Second, the matched Qwen 5-Hz and 1-Hz cached rows test whether an advantage is
explained by more frequent causal state refresh rather than the proposed local
evidence route. For native 1-Hz streams, the state emitted at the end of each
one-second chunk is never backfilled into earlier 200-ms rows; source gaps and
session boundaries reset all modality and language-model caches. MiniCPM-o is
a secondary cross-family diagnostic under the same 1-Hz schedule, not a
same-frequency architecture ranking. Its 5-Hz cached-branch row is omitted if
cached/full-prefix equivalence, cutoff invariance, or reset audits fail.

\todo{Freeze native chunk boundaries, initialization of rows before the first
complete chunk, cache-reset rules, the cached/full-prefix tolerance, and the
future-media intervention test. Report within-backbone 5-Hz versus 1-Hz
deltas plus cache/no-cache throughput and memory; do not interpret unmatched
Qwen--MiniCPM raw scores as a pure backbone effect.}

\subsection{Limitations and Qualitative Analysis}

The results support \todo{precise supported claim}, but several limitations
remain.
First, \todo{EasyCom size/session diversity limitation}.
Second, the adapter relies on \todo{state the actual visible-person box or
tracking assumption}.
Third, \todo{failure condition: occlusion, overlapping speakers, rapid floor
exchange, etc.} remains challenging.

\todo{Add one compact, pre-selected qualitative timeline only if space permits:
one multi-other false-entry improvement paired with one preserved real wearer
onset or one failure. Examples illustrate aggregate rates and never establish
the claim by themselves.}

\section{Conclusion}
\label{sec:conclusion}

This study asks whether a general-purpose Omni representation can be adapted
for causal floor forecasting of an invisible egocentric wearer. The candidate
method separates long-context Omni understanding from short-window local floor
evidence without assuming persistent speaker identity. \todo{State the final
learned-method improvement, matched-recall entry/INT false-positive rates,
fixed-horizon event AP, causal-control evidence, and main limitation. If no learned
variant passes these tests, reframe the submission as an information audit or
do not make the method claim.}

\appendix
\section{LLM-Assisted Behavioral Event Annotation}
\label{app:event_annotation}

EasyCom's raw speech boundaries are acoustic annotations rather than floor
labels. We preserve them unchanged for dense forecasting and create a separate
behavioral-event layer. Its design principle is to anchor an observed wearer
behavior first and then retrieve the conversational opportunity that could make
that behavior meaningful. False-positive opportunities are mined independently
because, by definition, many contain no wearer transition.

\subsection{Positive-event mining}

\textbf{Wearer activity transitions.}
We preserve every raw wearer $0\!\rightarrow\!1$ and $1\!\rightarrow\!0$
boundary as a semantic review candidate. Candidate construction does not merge
short silence gaps, infer within-turn resumption, or decide whether overlapping
speech constitutes takeover. It only records participant-resolved temporal and
transcript facts. The source intervals and dense targets remain unchanged.

\textbf{Wearer entry.}
Every raw wearer onset at $t_b$ is proposed as one wearer-entry review item.
We search backward within the retrieval window for non-wearer offsets and store
each one with its response delay and factual indicators of intervening wearer
speech, intervening other onsets, and concurrent activity. These are possible
context anchors rather than deterministically admissible causes. The LLM and
human reviewer decide whether the boundary is a genuine new entry, a
within-turn resumption, a wearer backchannel, or an excluded non-event. A
preceding release is not required for a positive entry: the wearer may enter
after a release, from silence, or while another speaker remains active. When a
release clearly prompts the entry, the LLM or human may select one supplied
option; neither may invent or move a timestamp.

\textbf{INT-yield.}
Every raw wearer offset $t_y$ is proposed as an INT-yield review item. The
generator records the current wearer activity-run start, all nearby non-wearer
onsets and offsets, activity immediately before and after $t_y$, whether an
other speaker is pre-existing or begins during wearer speech, strict and
50-ms-tolerant boundary crossing, overlap duration, the next speech onset,
transcripts, and target-of-speech metadata. Relations such as
$t_w<t_o<t_y<t_o^{\mathrm{off}}$ are descriptive evidence only: they neither
admit nor reject a candidate and do not establish takeover. The LLM and human
annotator answer whether the wearer stopped because a non-wearer genuinely took
over the floor. Natural completion, post-completion activity, and within-turn
pauses are therefore resolved semantically rather than pre-filtered.
When a wearer activity run is itself only a backchannel within a non-wearer's
retained turn, its onset supplies the false-entry opportunity; its trailing
offset is excluded rather than counted again as a false-yield opportunity.

\subsection{False-positive opportunity mining}

Hard-negative regions are retrieved independently of positive wearer
transitions. For entry, a same-speaker pause followed by resumption and an
other-to-other handover while the wearer stays silent are false-entry
opportunities. For INT, a short non-wearer backchannel or a non-takeover overlap
during which the wearer retains and continues the floor is a false-yield
opportunity. The timing rules provide candidate types, not final semantic
labels.

\subsection{Semantic pre-labeling and human adjudication}

The annotation assistant jointly receives immutable timestamps,
participant-resolved activity, neighboring boundaries, gap and overlap
measurements, transcripts, and target-of-speech metadata when available. It is
instructed not to equate raw onset with floor entry, offset with yield, or silence
with an entry opportunity,
or overlap with interruption, and not to invent unavailable prosody, gaze, or
intent. EasyCom's \texttt{[L]} marker supplies laughter evidence only and is not
by itself a sustained floor take or takeover. Preceding release options are
optional metadata for a positive wearer-entry pre-label.

We use \texttt{deepseek-v4-flash} at temperature zero with a primary and a
skeptical pass. Disagreement on verdict, event type, or hard-negative subtype
produces an \textsc{Exclude} suggestion. Every candidate then receives mandatory
human adjudication: accept the complete pre-label and optional release metadata,
relabel it, or exclude an
unscorable region. Only human-confirmed \textsc{Wearer-Entry},
\textsc{INT-Yield}, and hard-negative opportunities enter evaluation;
\textsc{Exclude} candidates do not. Thus \textsc{Exclude} is the sole
non-scoring verdict, but every exclusion retains a closed-vocabulary reason and
derived aggregation group. We report valid-but-out-of-scope behavior (e.g.,
natural wearer completion coincident with feedback), insufficient semantic
evidence, and data-quality failures separately; model-pass disagreement belongs
to insufficient evidence and is never counted as corrupt data.
Annotation-only transcripts, identities, target-of-speech metadata, LLM output,
and future observations are never supplied to the evaluated forecasting model.

\todo{Before freezing this annotation protocol, audit randomly sampled
continuous timeline segments without candidate hints to estimate candidate
recall for entry, INT, and each hard-negative family. Independently double-label
a pre-registered subset, report its size and agreement statistic, freeze all
candidate parameters and prompt/model identifiers on Sessions 4--12, and
report final candidate, control, exclusion, and adjudicated-label counts.}

\bibliographystyle{IEEEbib}
\bibliography{references}

\end{document}