Corvid: Bioacoustics, AI, and Re-ID

• Corvid is a multifaceted term spanning avian bioacoustics, multimodal chain-of-thought AI, and ecological video re-ID for bird identification.
• Bioacoustic studies reveal that corvid calls encode individual identity via rich, dynamic multi-harmonic structures analyzed with adaptive spectral methods and LPC.
• Computational systems 'Corvid' and 'CORVID' leverage hybrid vision encoders and precise segmentation techniques to enhance multimodal reasoning and wildlife tracking.

Corvid denotes, in the research literature represented here, both a biological referent and two computational systems. In the biological sense, it refers to the scientifically important corvid family—jackdaws, crows, ravens, and related birds—whose calls can exhibit rich dynamic multi-harmonic structure and encode individual identity (Stowell et al., 2016). In recent machine-learning literature, "Corvid" names a multimodal LLM with enhanced chain-of-thought reasoning capabilities (Jiang et al., 10 Jul 2025), while "CORVID" names a COlouR-based Video re-ID pipeline for assigning individual identities to ring-marked wild birds in the CHIRP dataset (Chan et al., 26 Mar 2026). The term therefore spans bioacoustics, multimodal reasoning, and ecological computer vision.

1. Scope of the term in current research

The term has three distinct but thematically related uses in the cited literature: a zoological one, a model name in multimodal AI, and an acronym for wildlife re-identification.

Usage	Definition	Source
corvid	The scientifically important corvid family, including jackdaws, crows, ravens, etc.	(Stowell et al., 2016)
Corvid	An MLLM with enhanced chain-of-thought reasoning capabilities	(Jiang et al., 10 Jul 2025)
CORVID	COlouR-based Video re-IDentification for ring-marked birds	(Chan et al., 26 Mar 2026)

The shared label is not merely nominal. In the biological literature, corvids are a model system for studying socially meaningful acoustic signals. In the computational literature, the reuse of the term for two AI systems situates it in multimodal reasoning and biodiversity monitoring. This suggests that "Corvid" functions as a cross-domain label linking complex animal communication, multimodal representation learning, and ecological measurement.

2. Corvid vocal complexity and the problem of individual identity

In bioacoustics, the corvid family is notable for calls that range from simple tones to rich dynamic multi-harmonic structures. The jackdaw material analyzed in the 2016 study exhibits multi-harmonic structure arising from two syrinx sources used simultaneously and rapid, dynamic pitch modulations—sweeps and sub-harmonics—over a wide frequency range of approximately $0.5$–$8$ kHz (Stowell et al., 2016). These properties make the signals poorly understood relative to simpler avian vocalizations.

The biological motivation is explicitly social. Jackdaws ($Corvus\ monedula$) live in socially complex groups and form lifelong pair bonds, and behavioural work shows that they recognise individual conspecifics by their contact calls (Stowell et al., 2016). The central research question is therefore not whether individuality is recoverable from calls, but where in the signal identity is encoded.

The standard short-time Fourier transform provides a baseline representation,

$$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$

but the cited study emphasizes two limitations for corvid calls: a fixed frequency grid cannot track rapid, nonlinear $f_0$ variations, and the representation cannot separate source and filter components (Stowell et al., 2016). In this setting, the methodological problem is tightly coupled to the biological one: the representational choice determines which putative identity cues become visible.

A plausible implication is that the analysis of corvid calls requires signal models that preserve onset dynamics, harmonic structure, and source-related variation simultaneously, rather than relying on generic spectrographic summaries.

3. Locating identity information in jackdaw calls

The 2016 study addresses the identity-localization problem by combining linear predictive coding, adaptive discrete Fourier transform, nearest-neighbour classification, and metric learning (Stowell et al., 2016). LPC models each sample as an autoregressive process of order $p$,

$x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$

with coefficient estimation defined by

$$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$

This yields both an LPC filter, $A(z)=1-\sum_k a_k z^{-k}$, interpreted as a vocal-tract approximation, and an LPC residual $e[n]$, interpreted as a syringeal source approximation (Stowell et al., 2016).

In practice, calls were processed by computing the STFT of raw $8$0, the STFT of the residual $8$1, and the LPC spectrum $8$2 as an alternate filter spectrogram. The adaptive discrete Fourier transform instead tracks a time-varying fundamental $8$3 rather than using a fixed frequency basis. Its algorithm initializes $8$4 in a broad range of $8$5–$8$6 Hz for songbirds, iteratively refines $8$7 to maximise harmonic peak energy, and reconstructs the spectrogram on a regular grid by nearest-neighbour resampling to match STFT resolution (Stowell et al., 2016). The stated advantages are that harmonics remain sharply localised under rapid pitch modulation and that sub- and inter-harmonics can be incorporated by allowing fractional multiples of $8$8.

Evaluation used a 3-nearest neighbours classifier with flattened spectrogram pixels or log-magnitudes, onset alignment of $8$9 ms to allow small temporal shifts, and both Euclidean $Corvus\ monedula$0 and Manhattan $Corvus\ monedula$1 distances (Stowell et al., 2016). Across 20 individuals, with chance at $Corvus\ monedula$2, raw-audio STFT achieved approximately $Corvus\ monedula$3 accuracy, STFT of the LPC residual approximately $Corvus\ monedula$4, and aDFT—refined or unrefined—also approximately $Corvus\ monedula$5. Combining LPC and aDFT yielded only marginal additional gain, and Manhattan distance consistently outperformed Euclidean distance (Stowell et al., 2016).

To identify the most informative time-frequency regions, the study used Large-Margin Nearest Neighbours and learned a Mahalanobis distance

$Corvus\ monedula$6

After training, feature-importance weights were computed as $Corvus\ monedula$7 and mapped back onto spectrogram pixels as a saliency heatmap (Stowell et al., 2016). The strongest saliency occurred immediately following call onset, in the first approximately $Corvus\ monedula$8–$Corvus\ monedula$9 ms, with a frequency region around $$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$0–$$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$1 kHz covering $$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$2 and the first two harmonics. The LPC residual further highlighted higher harmonics near onset, suggesting syringeal-source cues, and these patterns held under both raw and LPC-preprocessed representations (Stowell et al., 2016).

The resulting interpretation is specific: individual identity cues are concentrated in the early portion of jackdaw contact calls; both the fundamental frequency sweep near $$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$3 kHz and its first harmonics near $$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$4–$$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$5 kHz carry stable, individual-specific signatures; and source-related features are sufficient for high-accuracy identification (Stowell et al., 2016). The study accordingly proposes behavioural playback manipulations of early call segments and harmonic content, physiological recordings focused on onset dynamics in the $$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$6–$$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$7 kHz band, and automated monitoring systems using LPC residual plus lightweight kNN or metric-learned embeddings.

4. Corvid as a multimodal LLM

In the 2025 machine-learning literature, Corvid is an MLLM with enhanced chain-of-thought reasoning capabilities (Jiang et al., 10 Jul 2025). Architecturally, it augments a frozen Llama-3.2-8B-Instruct LLM with a hybrid vision encoder and a GateMixer connector. The hybrid encoder combines two off-the-shelf vision backbones operating on the same $$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$8 image: SigLIP ViT-S/400M and OpenCLIP ConvNeXt-XXL (Jiang et al., 10 Jul 2025).

The ViT-based encoder splits the image into $$X(t,f)=\sum_{n} x[n]\,w[n-t]\,e^{-j2\pi f n / f_s},$$9 patches, producing $f_0$0 embedded patches, processes them through $f_0$1 transformer blocks, and pools the final output into a $f_0$2-token spatial grid projected to dimension $f_0$3, giving $f_0$4 (Jiang et al., 10 Jul 2025). The CNN-based encoder likewise produces $f_0$5 spatial locations, each of dimension $f_0$6, yielding $f_0$7. These feature sets constitute the hybrid visual representation.

GateMixer aligns the visual tokens with the LLM embedding space. Each modality is linearly projected into a shared $f_0$8-dimensional space: $f_0$9 A gate is then computed from the concatenated features,

$p$0

and used for element-wise mixing,

$p$1

Finally, $p$2 learnable prefix embeddings are prepended and projected into the LLM token embedding space (Jiang et al., 10 Jul 2025). The connector is therefore არა a simple projector; it is a dynamic cross-modal alignment module with learnable context tokens.

Corvid’s training data and schedule are designed specifically for CoT behavior. MCoT-Instruct-287K begins from approximately $p$3K examples drawn from seven manually-crafted reasoning sets and three AI-assisted CoT generators. Manually created rationales are rewritten by a GPT-4o-based CoT Rewriter to expand, clarify, and standardize intermediate steps while preserving logical validity. All rewritten or AI-generated CoTs are then scored by GPT-4o on faithfulness, relevance, and completeness in $p$4, and only those with average score at least $p$5 are retained, yielding $p$6K single-turn $p$7 instances (Jiang et al., 10 Jul 2025).

Training proceeds in two stages. Stage I is multi-grained alignment pre-training on MAG-1M, approximately $p$8M image-text pairs of varying granularity, with LLM and vision backbones frozen and only GateMixer trained using caption-generation cross-entropy plus a bi-directional contrastive term (Jiang et al., 10 Jul 2025). Stage II is CoT-enhanced supervised fine-tuning on Corvid-1M, a $p$9M-example instruction set in which approximately $x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$0 of the data is reasoning and $x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$1 of that is CoT-formatted. The objective is standard autoregressive cross-entropy over both direct and step-by-step targets, after which the model becomes Corvid-base, an MLLM that can follow instructions and produce CoT traces (Jiang et al., 10 Jul 2025).

5. Inference-time scaling, self-verification, and benchmark profile

Corvid supplements training-time CoT supervision with an inference-time scaling strategy intended to mitigate both under-reasoning and over-reasoning (Jiang et al., 10 Jul 2025). Two prompts are issued in parallel to the same model: a direct prompt $x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$2 and a CoT prompt $x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$3. Each branch produces a final answer, a cross-modal similarity score $x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$4, and a confidence estimate derived from normalized perplexity,

$x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$5

If the two branches agree on the answer, that answer is returned. Otherwise, the system computes

$x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$6

for each branch and selects the branch with higher $x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$7 (Jiang et al., 10 Jul 2025). The stated computational cost is only two forward passes, without beam-search-style path enumeration.

Empirically, the model is evaluated on 9 multimodal reasoning benchmarks and 4 comprehensive understanding tests. On science problem-solving tasks, Corvid-o1-8B reports $x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$8 on MMStar, $x[n]\approx\sum_{k=1}^p a_k\,x[n-k]+e[n],$9 on MMMU, $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$0 on SQA-IMG, and $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$1 on AI2D, with an average of approximately $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$2, surpassing LlamaV-o1 at $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$3 and LLaVA-o1 at $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$4 (Jiang et al., 10 Jul 2025). On mathematical reasoning, it reports $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$5 on MathVista, $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$6 on MathVerse, $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$7 on WeMath, $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$8 on MathVision, and $$\min_{a_1,\dots,a_p}\,\sum_{n=p+1}^N \Bigl(x[n]-\sum_{k=1}^p a_k\,x[n-k]\Bigr)^2.$$9 on DynaMath, with the paper noting gains of $A(z)=1-\sum_k a_k z^{-k}$0 to $A(z)=1-\sum_k a_k z^{-k}$1 points over non-CoT models. On comprehensive understanding, it reports $A(z)=1-\sum_k a_k z^{-k}$2 on SEED-IMG, $A(z)=1-\sum_k a_k z^{-k}$3 on MMT-Val, $A(z)=1-\sum_k a_k z^{-k}$4 on RWQA, and $A(z)=1-\sum_k a_k z^{-k}$5 on BLINK, averaging approximately $A(z)=1-\sum_k a_k z^{-k}$6 (Jiang et al., 10 Jul 2025).

The ablation profile attributes performance to each major design choice. GateMixer improves average performance by $A(z)=1-\sum_k a_k z^{-k}$7 relative to MLP connectors; Corvid-1M with standardized CoTs yields $A(z)=1-\sum_k a_k z^{-k}$8 versus $A(z)=1-\sum_k a_k z^{-k}$9 for raw data and $e[n]$0 for direct-only data; self-verification yields $e[n]$1 versus $e[n]$2 for direct-only and $e[n]$3 for CoT-only; and the hybrid encoder yields $e[n]$4 versus $e[n]$5 and $e[n]$6 for the single-backbone alternatives (Jiang et al., 10 Jul 2025). On MathVista, inference speed is reported as $e[n]$7 s per instance, compared with $e[n]$8 s for LLaVA-o1 and $e[n]$9 s for LlamaV-o1, while maintaining higher accuracy.

Within the stated evidence, Corvid is therefore positioned as an open-source MLLM designed to improve structured multimodal reasoning through architecture, curated CoT data, staged optimization, and lightweight self-verification rather than through large inference-time search.

6. CORVID for individual bird re-identification in the wild

The 2026 CHIRP work introduces CORVID—COlouR-based Video re-IDentification—as an end-to-end framework for assigning individual identities to short video clips of ring-marked Siberian jays (Chan et al., 26 Mar 2026). The pipeline exploits the fact that each jay carries a unique combination of up to four colored plastic leg rings. Input consists of a 1 s video clip of 25 frames containing a single jay, cropped around a YOLOv8 detection, together with a metadata list of possible birds present in the territory, typically $8$00–$8$01 individuals and sometimes known neighbours (Chan et al., 26 Mar 2026).

The first stage is instance segmentation of rings using Mask2Former, fine-tuned on a ring-segmentation dataset containing 944 images, 2,713 ring masks, and 12 color classes (Chan et al., 26 Mar 2026). The output is binary masks and pixel-precise instance contours for all detected rings in each frame. Each ring mask is then cropped to a $8$02 pixel patch, converted to HSV, and passed as a $8$03 tensor to a Random Forest classifier trained to predict one of 12 color classes. For each detected ring patch $8$04 in frame $8$05, the classifier outputs a probability distribution $8$06 over the ring colors (Chan et al., 26 Mar 2026).

Temporal pooling operates over ring pairs matched by nearest centroids within each frame according to top-bottom and left-right position from the bird’s perspective. For a candidate pair $8$07 and color pair $8$08, the frame-wise score is

$8$09

and the score across $8$10 frames is the sum of the frame-wise scores (Chan et al., 26 Mar 2026). Identity matching then uses a lookup database of each bird’s ring-color tuple $8$11, assigning the clip to the known bird in the possible metadata with the highest match score. The paper also presents this as a simplified maximum a posteriori decision with a uniform prior over possible individuals (Chan et al., 26 Mar 2026).

The pipeline is modular but fully quantified. The segmentation subnetwork is Mask2Former with a ResNet50 or similar backbone trained with cross-entropy mask loss and dice loss; the color classifier is a Random Forest with 100 trees on flattened HSV histogram features; post-processing includes morphological cleanup, median-vote over overlapping detections, and centroid-distance clustering for pairing (Chan et al., 26 Mar 2026). Application-specific metrics include proportion correct frame assignments, feeding-event precision, recall, and $8$12, feeding-rate error per individual, co-occurrence rate error, and population-level mean or median absolute error with Pearson’s $8$13.

Performance is reported at multiple levels. On held-out $8$14 of the mask dataset, the random forest achieves approximately $8$15 per-ring color accuracy (Chan et al., 26 Mar 2026). For video re-ID under a closed-set, within-territory constraint, CORVID reaches Top-1 $8$16 and Top-3 $8$17, compared with pre-trained MegaDescriptor at $8$18 and $8$19, and fine-tuned MegaDescriptor at $8$20 and $8$21. On a disjoint set matching across territories with data-split individuals, CORVID reaches Top-1 $8$22 and Top-3 $8$23, compared with $8$24 for pre-trained MegaDescriptor and $8$25 for fine-tuned MegaDescriptor (Chan et al., 26 Mar 2026).

When integrated into an application-specific pipeline combining YOLOv8 detection, BoTSORT tracking, ID via a re-id method, and C3D action recognition, CORVID yields an individual-ID proportion of correct frames of $8$26, compared with $8$27 for MegaDesc and $8$28 for random assignment (Chan et al., 26 Mar 2026). Feeding-event precision, recall, and $8$29 are $8$30 for CORVID, $8$31 for MegaDesc, and approximately $8$32 for random assignment. Mean individual feeding-rate error is $8$33 for CORVID versus $8$34 for MegaDesc, while co-occurrence rate error is $8$35 for CORVID versus $8$36 for MegaDesc (Chan et al., 26 Mar 2026). Run-time is reported as approximately $8$37 FPS for Mask2Former segmentation on a single NVIDIA RTX3060, $8$38 ring patches per second for Random Forest color classification on CPU, and approximately $8$39 clips per second for the full pipeline.

The limitations are explicit. CORVID depends on reliable ring segmentation, so occlusions or blurred rings reduce accuracy; the current identity-matching stage assumes a closed set of possible birds, leaving open-set identification unresolved; and ring combinations are assumed to remain unchanged, so extended or swapped rings require manual updating (Chan et al., 26 Mar 2026). Future work is proposed in three directions: integrating appearance-based embeddings such as Multi-descriptor to handle unknowns, improving detector and segmentation backbones for faster inference, and designing an end-to-end deep network for joint segmentation and identity assignment.