# Knowledge Distillation for Small AI Models

Knowledge distillation represents a pivotal methodology in artificial intelligence, addressing the fundamental tension between the increasing computational demands of frontier large language models (LLMs) and the strict resource constraints of real-world deployment environments. As deep learning architectures scale to trillions of parameters, their utility in edge devices, real-time applications, and cost-sensitive enterprise pipelines becomes highly restricted [cite: 1, 2]. Knowledge distillation operates as a sophisticated compression and transfer learning paradigm wherein the behavior, feature representations, and reasoning capacities of a massive, highly capable "teacher" model are mathematically transferred to a significantly smaller, efficient "student" model [cite: 1, 3, 4]. 

The concept extends far beyond mere file compression; it is an epistemological mechanism by which an over-parameterized neural network, having mapped the complexities of vast training corpora, acts as an algorithmic pedagogue. This process democratizes access to state-of-the-art natural language processing (NLP) and reasoning capabilities, enabling local execution on consumer hardware, reducing carbon footprints, and lowering inference latency [cite: 5, 6, 7].

## Fundamental Mechanisms of Knowledge Transfer

The conceptual origin of transferring knowledge between neural networks dates back to model compression research by Rich Caruana in 2006, but the modern paradigm of knowledge distillation was formalized by Geoffrey Hinton, Oriol Vinyals, and Jeff Dean in 2015 [cite: 4, 8, 9]. The core premise is that large, over-parameterized models, or ensembles of models, learn complex generalizations that are not strictly necessary for final representation, but are essential for the optimization and feature discovery phases of training [cite: 8, 10]. Once these robust generalizations are learned, they can be taught directly to a much smaller model.

### Soft Targets and Dark Knowledge

In standard supervised machine learning, models are trained on "hard targets," which are typically one-hot encoded vectors where the correct class is assigned a probability of 1 and all incorrect classes are assigned a 0. However, a well-trained teacher model outputs "soft targets"—a continuous probability distribution across all possible classes [cite: 4, 8]. These soft labels contain what researchers term "dark knowledge." 

For example, an image classification model might predict a 95% probability that an image contains a cat, a 3% probability it is a leopard, and a 1% probability it is a dog [cite: 9]. The non-zero probabilities assigned to incorrect classes reveal the teacher's internal understanding of semantic similarity, structural relationships, and feature overlap. By forcing the student model to learn this nuanced probability distribution rather than just the absolute right answer, the student absorbs the teacher's generalized logic [cite: 9].

### Mathematical Formulation and Temperature Scaling

To effectively extract this knowledge, distillation modifies the softmax activation function used in the output layer of the neural network. The standard softmax calculates the probability $p_i$ of a class $z_i$ as:

$$p_i = \frac{\exp(z_i)}{\sum_j \exp(z_j)}$$

Knowledge distillation introduces a scalar hyperparameter called "temperature" ($T$). Both the teacher and the student apply this temperature to their pre-softmax logits during the distillation training phase:

$$p_i = \frac{\exp(z_i / T)}{\sum_j \exp(z_j / T)}$$

Applying a higher temperature ($T > 1$) forces the resulting distribution to become softer and more uniform [cite: 11]. This amplification elevates the probabilities of near-zero classes, preventing them from being mathematically ignored and forcing the student model to learn the structural relationships between unlikely output categories [cite: 11]. The student is then trained using a composite, multi-term loss function. This function typically calculates a weighted combination of the Kullback-Leibler (KL) divergence between the student's and teacher's temperature-scaled soft predictions, alongside the standard cross-entropy loss calculated against the ground-truth hard labels without temperature scaling [cite: 12, 13, 14].

[image delta #1, 0 bytes]





## Synergies With Hardware and Model Compression

Knowledge distillation is rarely deployed in commercial isolation. To achieve maximal efficiency for edge inference and IoT applications, distillation forms the architectural backbone of a broader, synergistic model optimization strategy that includes quantization and pruning [cite: 2, 15]. While distillation targets the transfer of algorithmic logic and behavior, quantization and pruning strictly target the physical storage and computational mechanisms of the artificial neural network [cite: 2, 15].

### Quantization Methodologies

Quantization shrinks the numerical precision of a model’s weights and internal activations [cite: 3, 15]. Default deep learning architectures typically represent parameters using 32-bit floating-point numbers (FP32). Quantization algorithms mathematically map these values to lower-precision formats, such as 16-bit floating-point (FP16), 8-bit integers (INT8), or lower [cite: 2, 15]. This reduction logarithmically shrinks the model's physical memory footprint on the hard disk and substantially accelerates tensor multiplication operations, particularly when executing on specialized hardware accelerators like INT8-capable GPUs or mobile Neural Processing Units (NPUs) [cite: 3, 15].

Quantization application generally takes two dominant forms in the industry: Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT). PTQ applies scale and zero-point mapping functions to a model that is already fully trained. While PTQ is computationally inexpensive and fast to apply, it can introduce severe rounding errors that degrade predictive performance [cite: 2]. QAT addresses this by simulating these low-precision rounding errors during the active training phase itself. By forcing the network to adjust its weights dynamically to account for the lowered precision before training concludes, QAT results in significantly higher fidelity and robustness upon deployment [cite: 2, 16].

### Structural and Magnitude-Based Pruning

Pruning aims to identify and systematically remove redundant, dormant, or non-contributing parameters within the neural network [cite: 2, 3, 15]. Magnitude-based pruning is the most ubiquitous approach; it evaluates the absolute numerical value of individual weights and severs connections that fall below a predefined threshold, operating on the assumption that near-zero weights exert minimal impact on the final output matrix [cite: 15].

Engineers categorize pruning into unstructured and structured methods. Unstructured pruning removes individual weights indiscriminately, creating highly sparse matrices. While theoretically efficient in reducing operations, standard hardware processors (like consumer CPUs and GPUs) struggle to process irregular sparse matrices efficiently without highly specialized computational kernels [cite: 15, 16]. Structured pruning, conversely, targets entire architectural components—such as specific neurons, individual attention heads, or even entire transformer layers (depth-pruning). This results in a denser, hardware-friendly model architecture that natively executes faster and draws less power on standard commercial processors [cite: 2, 15].

### Joint Optimization Pipelines

The most performant lightweight models are created through unified pipelines that execute these techniques concurrently. Frameworks such as the Joint Pruning, Quantization, and Distillation (JPQD) approach, utilized within the OpenVINO Neural Network Compression Framework, optimize the transformer model simultaneously [cite: 2, 16]. In this sequence, a large teacher is first structurally pruned to remove architectural bloat. The now-sparse teacher is then utilized to distill its knowledge into a specialized student model. Finally, the student model undergoes Quantization-Aware Training (QAT) to map its remaining operations to low-precision hardware instructions [cite: 2, 16]. This multiplicative, parallel optimization approach recovers the accuracy lost during pruning and quantization far more effectively than isolated, sequential fine-tuning [cite: 2, 16].

## Systems Engineering and Distributed Training Bottlenecks

Moving knowledge distillation for large language models from theoretical frameworks to production-scale infrastructure reveals massive systems-level bottlenecks. When distilling frontier LLMs comprising hundreds of billions of parameters, the primary constraints immediately shift from pure model architecture to the limitations of distributed computing networks and data transfer protocols [cite: 17].

### Architectural Mismatch and Compute Profiles

A frequent and catastrophic architectural error in modern distillation pipelines involves attempting to execute both the massive teacher and the learning student model on a homogeneous computing backend [cite: 17]. The computational workload profiles of the two models are diametrically opposed. 

The teacher model requires high-throughput forward inference. It is heavily memory-bandwidth bound and is highly reliant on maximizing large Key-Value (KV) caches and leveraging specialized inference kernels like vLLM or TensorRT-LLM to generate tokens rapidly [cite: 17]. Conversely, the student model requires high-throughput backward passes to update its weights during training. The student is heavily compute-bound, requiring complex optimizer state management, gradient accumulation mechanisms, and frameworks like Fully Sharded Data Parallel (FSDP) to manage its internal updates [cite: 17]. Treating them as a single computational unit leads to severe hardware underutilization. Optimization requires decoupling them into separate operating processes, or entirely separate hardware clusters, allowing the teacher to scale as a remote inference service independently of the student [cite: 17].

### The Logit Tax and Network Saturation

Once the models are decoupled, the physical network connecting the inference cluster and the training cluster becomes the primary bottleneck. Transmitting the full probability distribution (the logits) from the teacher to the student across network nodes—dubbed the "Logit Tax"—is prohibitively expensive in terms of bandwidth [cite: 17]. 

In a standard LLM distillation configuration utilizing a conservative batch size of 8, a sequence length of 2048 tokens, a standard vocabulary size of 150,000 (typical of models like Qwen or Gemma), and 2-byte precision (bfloat16), shipping the full logit tensor requires approximately 15 gigabytes of data transfer per single training step [cite: 17]. When utilizing an ensemble of teachers, this data volume easily saturates even high-end, specialized Network Interface Cards (NICs), starving the student model of labels and halting training [cite: 17]. Distributed systems engineers mitigate this severe bottleneck by transmitting the much narrower hidden states instead of full logits across the network. A specialized shim layer on the student node then recomputes the logits locally using the teacher's shared Language Model (LM) head. If the architectures diverge, a learned linear projection layer maps the student's hidden dimension to the teacher's logit space, effectively eliminating the network bandwidth constraint [cite: 17].

### The Model Capacity Gap

A critical algorithmic bottleneck inherent to the methodology is the "capacity gap." If the teacher model is vastly more complex or parameterized than the student model, the student structurally lacks the representational capacity to mimic the teacher's highly sophisticated, multi-modal probability distributions [cite: 1, 18]. This severe architectural mismatch can lead to ineffective knowledge transfer, mode collapse, and overall predictive performance degradation, negating the purpose of distillation [cite: 1, 18].

### Intermediate Curriculums and Difficulty-Aware Filtering

To bridge this capacity gap, researchers employ intermediate models and smoothing techniques. The Tutor-Enhanced Iterative Distillation (TEID) framework introduces an intermediate-sized "tutor" model. The teacher distills knowledge to the tutor, which then distills knowledge to the final student, breaking the learning curve into manageable gradients [cite: 1]. Similarly, Skew Kullback-Leibler (SKL) divergence methods introduce an intermediate target distribution that interpolates directly between the teacher's and the student's outputs, mitigating distributional divergence and preventing the student network from being overwhelmed by complexity it cannot map [cite: 19].

Standard distillation processes also frequently waste massive amounts of computational time by forcing the student to process and learn from simple data samples it already fully comprehends [cite: 20]. Difficulty-Aware Knowledge Distillation (DA-KD) frameworks dynamically adjust the distillation dataset based on the student's evolving competency [cite: 1, 20]. By employing routing mechanisms or difficulty graders aligned with human benchmarks, DA-KD focuses the loss functions entirely on hard samples that the student currently fails [cite: 20]. This intelligent filtering reduces total training costs by half while achieving equivalent or superior benchmark performance [cite: 20]. To prevent gradient explosion or vanishing gradient problems when the dataset becomes heavily skewed toward exceptionally difficult samples, stabilizing mechanisms like Bidirectional Discrepancy Loss (BDL) are applied to smooth the optimization trajectory [cite: 20].

## Distillation of Complex Reasoning and Chain-of-Thought

The advent of highly capable LLMs empirically demonstrated that complex mathematical and logical reasoning is an emergent property tightly correlated with massive parameter counts [cite: 21, 22]. Techniques like Chain-of-Thought (CoT) prompting—where models are explicitly primed to verbalize intermediate rationalizations step-by-step prior to delivering an answer—yield dramatic performance gains. However, historical scaling laws indicated these benefits generally only emerged for models exceeding 50 billion parameters [cite: 22]. Small language models naturally struggle to generate accurate reasoning traces independently, often producing flawed logic, hallucinating facts, or entirely skipping critical inferential steps [cite: 21, 23]. 

### Symbolic and Adaptive Distillation Approaches

To instill advanced reasoning in SLMs, researchers utilize Symbolic Chain-of-Thought Distillation (SCoTD). Instead of relying strictly on white-box logit transfer, SCoTD operates as a data-centric distillation pipeline. It samples vast corpora of text containing high-quality, step-by-step reasoning generated by a massive teacher model [cite: 22]. The student is then trained on these rationalizations via supervised fine-tuning (SFT). This approach empirically proved that models as small as 125 million to 1.3 billion parameters can learn to "think" step-by-step, provided they are exposed to a sufficiently diverse and rigorously filtered set of reasoning chains [cite: 22].

However, simplistic SFT on massive reasoning traces often produces overly verbose student models that babble unnecessarily, leading to severe inefficiencies at inference time due to excessive token generation [cite: 24]. Difficulty-Aware CoT Distillation (DA-CoTD) solves this by adapting the length and complexity of the reasoning trace directly to the inherent complexity of the user's input prompt. DA-CoTD utilizes an LLM-based grader to compress verbose teacher traces into highly efficient, difficulty-aligned chains before feeding them to the student via Direct Preference Optimization (DPO), maintaining accuracy while reducing inference token generation by up to 30% [cite: 24].

### Mistake-Driven Learning and Dual Pathways

Standard CoT distillation methodologies typically train student models exclusively on the correct outputs generated by the teacher [cite: 23]. However, empirical analysis of reasoning traces shows that they consist mostly of simple syntactic formatting, with only a tiny fraction (approximately 4.7%) representing the "key reasoning steps" that actually dictate the final logical conclusion [cite: 23]. When trained only on positive, correct samples, students often merely imitate the structural form and vocabulary of the reasoning without actually learning the pivotal logic steps required to solve novel problems [cite: 23].

The EDIT (mistakE-Driven key reasonIng step distillaTion) framework models human pedagogical methods by providing the student with dual reasoning paths [cite: 23]. Through specialized prompting, the teacher model is forced to generate two CoT outputs for a single problem: one that successfully reaches the correct answer, and one that fails due to a specific induced logical error [cite: 23]. By applying a minimum edit distance algorithm between the two contrasting paths, the framework isolates the exact logical juncture where the failure occurred. Utilizing these negative samples in the distillation loss function forces the student to recognize and internalize the specific step that dictates accuracy, moving the student beyond mere stylistic imitation to actual capability acquisition [cite: 21, 23].

### Implicit Chain-of-Thought

A novel and highly experimental frontier in reasoning distillation aims to entirely internalize the verbalized steps of CoT. Implicit Chain-of-Thought reasoning attempts to conduct reasoning "vertically" across the model's internal hidden layers rather than "horizontally" by outputting intermediate text tokens into the context window [cite: 25]. In this method, a teacher model explicitly generates a CoT sequence, and its internal hidden states (across all layers and generated tokens) are extracted. An emulator network is then trained to mathematically predict these multi-layer state matrices. This allows the student model to process the logical progression entirely within its hidden states, vastly accelerating inference times by bypassing the sequential generation of intermediate words [cite: 25].

## Evolution of Distilled Natural Language Understanding

The history of model distillation can be traced precisely through the evolution of foundational encoder models to the current era of massively distilled, autoregressive decoder-only models. 

### The BERT Era Compression

Following the release of Google's Bidirectional Encoder Representations from Transformers (BERT), the natural language processing community faced acute latency challenges due to the model's 110-million parameter footprint, making it difficult to deploy on mobile applications or edge servers [cite: 7, 26]. Hugging Face introduced DistilBERT, pioneering the practical application of KD for transformers [cite: 10, 27]. By removing token-type embeddings and the pooler, and reducing the layer count by half, DistilBERT achieved a 40% reduction in parameters and a 60% acceleration in inference time [cite: 10, 28]. Utilizing a triple loss function that combined masked language modeling loss, classical distillation loss, and cosine-distance loss to strictly align the student and teacher hidden vectors, DistilBERT maintained 97% of BERT's original performance on the ubiquitous GLUE benchmark [cite: 10, 28, 29].

Subsequent iterations pushed compression algorithms further. DistilRoBERTa implemented similar layer-halving techniques on the highly optimized RoBERTa architecture, achieving 2x faster inference and near-parity on downstream sentiment and cybersecurity analysis tasks [cite: 14, 26, 30]. TinyBERT represented an aggressive, generalized layer-wise distillation strategy that mapped every third BERT layer to the student. This reduced BERT-base to just 14.5 million parameters—a staggering 7.5x reduction in size and 9.4x acceleration in speed—while still achieving a highly competitive 77.0% GLUE score compared to BERT's 79.5% [cite: 31, 32, 33].

[image delta #2, 0 bytes]





### The Open-Weights Era and Massive Language Models

The introduction of the Llama series by Meta shifted the industry focus from efficient encoders to massive decoder-only generative models. The Llama 3 family includes 405B, 70B, and 8B parameter variants [cite: 34, 35]. While the 405B model serves as the flagship, setting state-of-the-art benchmarks for open-weights, its massive computational overhead restricts its deployment [cite: 35, 36]. Meta utilized the 405B model extensively to generate, curate, and filter the synthetic training data for the 70B and 8B variants, demonstrating a form of data-pipeline distillation that embedded high-order reasoning into the smaller architectures [cite: 37]. 

The resulting performance gap highlights the efficacy of this approach. The 70B model achieves an MMLU (Massive Multitask Language Understanding) score of ~86.0% and a GSM8K (Grade School Math) score of 95.1%. The highly compressed 8B parameter model achieves an MMLU score of ~73.0% and a GSM8K score of 84.5%—scores that comfortably eclipsed previous generation models nearly ten times its size, making it a staple for on-device deployment [cite: 35, 38, 39].

| Meta Llama 3.1 Model Tier | MMLU (Reasoning & Knowledge) | GSM8K (Math & Logic) | HumanEval (Coding) | ARC-Challenge |
| :--- | :---: | :---: | :---: | :---: |
| Llama 3.1 70B | ~86.0% | ~95.1% | ~80.5% | ~94.8% |
| Llama 3.1 8B | ~73.0% | ~84.5% | ~72.6% | ~83.4% |

Third-party projects have pushed these architectures even further to target the strict constraints of the edge-deployment market. The TinyLlama project utilized the exact architecture and tokenizer of Llama 2 but trained a 1.1 billion parameter model from scratch on a highly curated synthetic and public dataset of 3 trillion tokens (sourced primarily from Slimpajama and Starcoderdata) [cite: 40, 41]. Executed over 90 days on just 16 A100-40G GPUs, the project proved that extreme data density could compensate for low parameter counts [cite: 40, 42]. Despite its diminutive size, TinyLlama outperforms comparable models like Pythia and OPT in zero-shot problem-solving tasks, demonstrating the efficacy of massive token exposure [cite: 43]. 

### Hardware Constraints and Local Execution Realities

The theoretical performance of these models often meets stark physical limitations during local deployment. For instance, execution on consumer hardware, such as the Apple Mac mini equipped with M4 Pro chips, reveals critical memory bandwidth bottlenecks [cite: 44]. While 3B to 8B parameter models execute flawlessly, attempts to run 70B-class models on local 32GB or 64GB unified memory architectures result in sluggish generation speeds of 3 to 5 tokens per second [cite: 44]. 

Furthermore, while SLMs (Small Language Models) under 8B parameters approach LLM performance on text classification and routing tasks, they often suffer severe regressions in structured output reliability [cite: 45, 46]. Benchmark evaluations indicate that while a model like Llama 3.1 8B executes JSON parsing and schema compliance reliably, scaling down to the 3B parameter tier results in parse rates dropping precipitously to roughly 50%, indicating that a minimum parameter density is required for strict syntactical adherence [cite: 45].

## DeepSeek-R1 and the Reinforcement Learning Distillation Paradigm

A definitive paradigm shift in reasoning distillation was demonstrated by DeepSeek with the release of the DeepSeek-R1 models in early 2025. Prior to this, achieving advanced mathematical and coding reasoning typically required computationally exhaustive Reinforcement Learning (RL) pipelines applied directly to the target model [cite: 47]. 

Rather than utilizing RL to train small models from scratch, DeepSeek extracted approximately 800,000 high-quality reasoning and non-reasoning data samples from their massive, RL-trained flagship model [cite: 47]. They utilized this highly curated dataset to conduct pure Supervised Fine-Tuning (SFT) on existing open-source base models, specifically the Qwen (1.5B to 32B) and Llama (8B and 70B) series, entirely bypassing the RL stage for the students [cite: 47].

The benchmark results proved highly disruptive to the industry standard. The DeepSeek-R1-Distill-Qwen-1.5B model achieved 28.9% on AIME 2024 and 83.9% on MATH-500, outperforming massive proprietary models like GPT-4o and Claude-3.5-Sonnet on specific mathematical benchmarks [cite: 47]. The 32B distilled variant achieved 72.6% on AIME and 94.3% on MATH-500, setting new global records for dense models and significantly exceeding the performance of specialized reasoning models like OpenAI-o1-mini [cite: 47].

| DeepSeek-R1 Distilled Model | Base Architecture | AIME 2024 | MATH-500 | GPQA Diamond | LiveCodeBench |
| :--- | :--- | :---: | :---: | :---: | :---: |
| DeepSeek-R1-Distill-1.5B | Qwen 2.5 | 28.9% | 83.9% | 33.8% | 16.9% |
| DeepSeek-R1-Distill-7B | Qwen 2.5 | 55.5% | 92.8% | 49.1% | 37.6% |
| DeepSeek-R1-Distill-8B | Llama 3.1 | 50.4% | 89.1% | 49.0% | 39.6% |
| DeepSeek-R1-Distill-14B | Qwen 2.5 | 69.7% | 93.9% | 59.1% | 53.1% |
| DeepSeek-R1-Distill-32B | Qwen 2.5 | 72.6% | 94.3% | 62.1% | 57.2% |
| DeepSeek-R1-Distill-70B | Llama 3.3 | 70.0% | 94.5% | 65.2% | 57.5% |

Crucially, the computational efficiency of this approach altered the economics of model training. The vanilla Llama 3.1 8B model required approximately 1.46 million H100 GPU-hours to pre-train [cite: 48]. Its distilled counterpart, DeepSeek-R1-Distill-Llama-8B, acquired superior reasoning capabilities across all benchmarks with only ~610 H100 GPU-hours—a compute expenditure reduction of over 99.9% [cite: 48]. 

## Generative Degradation and Model Collapse

While distillation methodologies yield exceptionally impressive headline benchmark scores, the fundamental nature of the algorithm introduces severe risks of long-term capability degradation. Knowledge distillation is inherently a lossy mathematical projection [cite: 49]. By optimizing a student model strictly to match a teacher’s output distribution or specific step-by-step reasoning traces, the student may perfectly mimic the primary observable metrics while fundamentally failing to preserve the underlying computational capabilities that make the teacher reliable in unpredictable edge cases [cite: 49].

Evaluations in the field often rely on the flawed assumption that retained task scores imply retained general capabilities [cite: 49]. However, researchers have identified that "off-metric" capabilities—such as strict adherence to safety guardrails, grounding in factual constraints, algorithmic process faithfulness, and overall generative diversity—frequently deteriorate drastically in distilled models [cite: 49]. A student model may generate code that passes a unit test (the benchmark), but it may achieve that result through brittle, memorized logic rather than generalized understanding.

This degradation becomes highly problematic and systemic when synthetic data generated via distillation is introduced back into the general pre-training corpora of future AI models. Iterative training loops, where subsequent generations of models continually learn from the synthetic, distilled outputs of their predecessors, trigger an exponential degradation phenomenon known as "model collapse" [cite: 50, 51]. As the recursive loop progresses, the minor statistical errors, hallucinations, and loss of variance introduced during the initial distillation phases become mathematically amplified. The resulting models rapidly lose their ability to model the true, underlying variance of human language and data distributions, eventually outputting homogenous, low-quality text [cite: 51, 52]. Mitigating this requires drawing on pedagogical frameworks, such as Vygotsky's Zone of Proximal Development (ZPD), to implement strict curriculum sequencing and mastery gating during synthetic data generation [cite: 50].

## Adversarial Distillation and Model Extraction Attacks

The massive operational efficiency and cost reductions provided by knowledge distillation have birthed a novel cyber-threat vector known as a "distillation attack" or "model extraction attack." In benign applications, developers use distillation to compress a model they own or have licensed for open use. In an adversarial distillation attack, an entity utilizes legitimate, authenticated access points (such as commercial APIs, mobile applications, or third-party resellers) to systematically and aggressively probe a proprietary frontier model [cite: 53, 54, 55]. 

The attacker harvests millions of high-value computational outputs—including complex reasoning traces, algorithmic code trajectories, formulated answers, and alignment preferences [cite: 54, 56]. The attacker then uses this massive synthetic dataset to conduct supervised fine-tuning on a localized open-source student model [cite: 54, 56].

### Capabilities Cloning and Safety Stripping

Distillation attacks allow adversaries to "free-ride" entirely on the massive capital expenditure invested by leading AI labs [cite: 57, 58]. By cloning the intelligence of a model without bearing the multi-million dollar data curation and pre-training compute costs, attackers can rapidly deploy competing products that undercut the original provider on service pricing [cite: 56]. 

Furthermore, because the attacker dictates the fine-tuning process of the student model, they possess the ability to intentionally and selectively distill raw intelligence capabilities while explicitly stripping away the teacher model's embedded safety protocols, ethical alignments, and ideological constraints [cite: 56, 57, 59]. The resulting adversarial model possesses frontier-level intelligence but is entirely devoid of the original developers' usage guardrails, posing severe risks for malicious exploitation [cite: 60].

### Geopolitical Ramifications and Trace Coercion

Distillation attacks gained international prominence during a high-profile controversy between US-based AI labs and the Chinese firm DeepSeek. OpenAI delivered a formal memorandum to the US Congress alleging that DeepSeek conducted an industrial-scale distillation attack, utilizing obfuscated third-party routers and programmable networks to bypass OpenAI's API access restrictions [cite: 57, 61, 62]. The allegations asserted that DeepSeek harvested vast quantities of outputs from frontier models like ChatGPT to artificially inflate the capabilities of their own local models at a fraction of standard training costs [cite: 61, 62].

Similar operational concerns were echoed by Anthropic and Google. Google's Threat Intelligence Group (GTIG) actively monitored and disrupted massive "reasoning trace coercion" campaigns [cite: 54, 55, 58]. In one documented instance, an attacker submitted over 100,000 highly targeted prompts specifically designed to force the Gemini model to output its normally hidden internal reasoning traces in non-English languages, representing a clear attempt to extract and clone the model's core cognitive architecture for foreign deployment [cite: 54, 55]. These incidents have fundamentally shifted the narrative around distillation from a benign academic engineering technique to a matter of critical national security, geopolitical competition, and intellectual property protection [cite: 58, 63].

## Legal Frameworks and Intellectual Property Friction

The surge in industrial-scale distillation attacks has exposed critical ambiguities in global intellectual property law regarding the protection of artificial intelligence capabilities and the enforceability of commercial protections [cite: 9, 63].

### Terms of Service Constraints and Enforceability

To protect their massive intellectual property investments, frontier AI developers uniformly deploy strict Terms of Service (ToS) explicitly prohibiting the use of their model outputs to train competing AI systems. OpenAI's policy explicitly dictates that users may not "Use Output to develop models that compete with OpenAI" [cite: 64]. Anthropic strongly restricts outputs from being used to "develop or train any artificial intelligence or machine learning algorithms or models" that constitute competitive products [cite: 65, 66]. Google’s Gemini API terms mandate similar prohibitions against reverse engineering or extracting parameter weights through generative queries [cite: 67, 68].

However, the legal enforcement mechanisms for these clauses are currently severely limited, usually resulting only in account termination and API bans [cite: 69]. Translating these ToS violations into actionable courtroom litigation is highly complex. The primary hurdle is the evidentiary difficulty of proving definitively that a specific competing model was explicitly trained on proprietary outputs, particularly when the competitor operates in a foreign jurisdiction with differing regulatory standards [cite: 69, 70]. 

### Copyright Law and the Human Authorship Requirement

The application of traditional copyright law to model distillation remains legally precarious and largely untested in high courts. In the United States, copyright protection strictly requires "human authorship" [cite: 9, 70]. The outputs generated by an AI API—which serve as the highly valuable training data for the student model during an attack—are the result of autonomous algorithmic computation, not human creative expression [cite: 9]. 

As evidenced by specific guidance from the US Copyright Office in cases like *Zarya of the Dawn*, AI-generated text, code, or images lack the human element required to be considered copyrightable works [cite: 9]. Therefore, an adversarial actor harvesting billions of tokens of AI output is arguably not infringing on any recognized copyright [cite: 69, 71]. Furthermore, because models like OpenAI's transfer output ownership to the user prompting the system, the platform developer surrenders the legal standing to claim infringement [cite: 70].

### Unfair Competition and Antitrust Perspectives

Given the severe limitations of copyright, legal scholars suggest that disputes over knowledge distillation will center on contract law and Unfair Competition frameworks [cite: 9, 70]. Plaintiffs will likely argue that distillation constitutes flagrant "free-riding," exploiting another entity's substantial R&D investment to create a direct market substitute in violation of fair commercial practices [cite: 9]. Alternatively, developers may pursue trade secret misappropriation claims, asserting that the model outputs contain derivative insights into the proprietary, confidential algorithms and datasets used to pre-train the teacher [cite: 69, 72].

Conversely, antitrust advocates and open-source proponents argue that broad "anti-distillation" clauses are inherently anti-competitive and potentially legally unenforceable [cite: 73]. From this doctrinal perspective, frontier models represent foundational infrastructure in the modern digital economy [cite: 73]. Restricting developers from extracting functional principles to build interoperable or competing systems may violate competition statutes. Scholars draw parallels to historical software reverse-engineering precedents, such as *Sega Enterprises v. Accolade*, where the extraction of functional principles was protected [cite: 73]. Under frameworks like the EU's Article 102 TFEU and the essential facilities doctrine, coordinated restrictions on output usage by dominant AI firms could face intense regulatory scrutiny [cite: 73]. As international regulatory bodies attempt to harmonize rules regarding AI IP protection, the legality of knowledge distillation remains a profound and unsettled friction point between open-source innovation and corporate intellectual property rights [cite: 63, 70].

## Sources

1. [Scaling LLM Distillation: Why Your Network is the Bottleneck (and How to Fix It)](https://medium.com/better-ml/scaling-llm-distillation-why-your-network-is-the-bottleneck-and-how-to-fix-it-ef2cf8822268)
2. [Mitigating the Capacity Gap in Knowledge Distillation](https://soe.lau.edu.lb/files/Tekli_conf_2025_MitigatingTheCapacityGap.pdf)
3. [Findings EMNLP 2025](https://aclanthology.org/2025.findings-emnlp.451.pdf)
4. [AAAI 2023 - Distilling Reasoning Capabilities](https://ojs.aaai.org/index.php/AAAI/article/view/29821/31426)
5. [ICML 2025 - Difficulty-Aware Knowledge Distillation](https://icml.cc/virtual/2025/poster/45516)
6. [How do various quantization methods compare](https://www.researchgate.net/post/How_do_various_quantization_methods_knowledge_distillation_frameworks_and_pruning_strategies_compare_when_reducing_the_size_of_AI_model_approaches)
7. [Joint Pruning, Quantization and Distillation](https://blog.openvino.ai/blog-posts/joint-pruning-quantization-and-distillation-for-efficient-inference-of-transformers)
8. [Efficient Inference at the Edge](https://uplatz.com/blog/efficient-inference-at-the-edge-a-comprehensive-analysis-of-quantization-pruning-and-knowledge-distillation-for-on-device-machine-learning/)
9. [Part 3 — Model Optimization Techniques](https://medium.com/@deep.bbd/part-3-model-optimization-techniques-a-deep-dive-into-quantization-pruning-and-knowledge-ef4a8c4d6d8b)
10. [Model Optimization Techniques](https://pub.aimind.so/model-optimization-techniques-pruning-quantization-knowledge-distillation-sparsity-2d95aa34ea05)
11. [US sounds alarm on China's AI distillation](https://asiatimes.com/2026/04/us-sounds-alarm-on-chinas-ai-distillation-as-deepseek-v4-debuts/)
12. [OpenAI Alleges China's DeepSeek Stole its Intellectual Property](https://www.fdd.org/analysis/2026/02/13/openai-alleges-chinas-deepseek-stole-its-intellectual-property-to-train-its-own-models/)
13. [Global Times - US tech firms clamp down](https://www.globaltimes.cn/page/202604/1358378.shtml)
14. [Legality of distillation](https://law.asia/openai-deepseek-ai-distillation/)
15. [Sam Altman's OpenAI accuses Chinese DeepSeek of copying AI models](https://timesofindia.indiatimes.com/technology/tech-news/sam-altmans-openai-accuses-chinese-deepseek-of-copying-ai-models-writes-letter-to-us-lawmakers/articleshow/128288882.cms)
16. [Hacker News - Anthropic ToS discussion](https://news.ycombinator.com/item?id=44429697)
17. [AI Terms of Service Fine Print](https://www.termsfeed.com/blog/ai-terms-service-fine-print/)
18. [OpenAI Terms of Use](https://openai.com/policies/row-terms-of-use/)
19. [ToS Watchdog - Anthropic](https://terms.law/ToS-Watchdog/ai-services/anthropic/)
20. [Can I use my outputs to train an AI model - Anthropic Support](https://support.claude.com/en/articles/12326764-can-i-use-my-outputs-to-train-an-ai-model)
21. [Adaptive Chain-of-Thought Distillation](https://www.mdpi.com/2227-7390/13/22/3646)
22. [Symbolic Chain-of-Thought Distillation](https://arxiv.org/html/2306.14050v2)
23. [Beyond Imitation: Learning Key Reasoning Steps from Dual CoTs](https://openreview.net/forum?id=aygBjpMdan)
24. [Implicit Chain of Thought Reasoning via Knowledge Distillation](https://openreview.net/forum?id=9cumTvvlHG)
25. [DA-CoTD: Efficient Chain-of-Thought Reasoning](https://neurips.cc/virtual/2025/130537)
26. [Benchmarked 12 small language models](https://www.distillabs.ai/blog/we-benchmarked-12-small-language-models-across-8-tasks-to-find-the-best-base-model-for-fine-tuning/)
27. [Comparative analysis of Llama 3-8B and DistilBERT](https://escholarship.org/content/qt8tr0p6c0/qt8tr0p6c0_noSplash_4a88210d64bd289d2826d86b255f6c98.pdf?t=skmg3p)
28. [Introduction to DistilBERT in Student Model](https://www.analyticsvidhya.com/blog/2022/11/introduction-to-distilbert-in-student-model/)
29. [Knowledge distillation for natural language processing](https://project-archive.inf.ed.ac.uk/ug4/20201880/ug4_proj.pdf)
30. [Response Selection for Long Reasoning Distillation](https://arxiv.org/html/2510.03988v1)
31. [Back to the Basic: Distilling the Knowledge in a Neural Network](https://medium.com/@nayounghoon0223/back-to-the-basic-a-foundational-knowledge-distillation-paper-of-distilling-the-knowledge-in-a-10fba70bab3c)
32. [Knowledge Distillation - NCBI](https://pmc.ncbi.nlm.nih.gov/articles/PMC8053015/)
33. [Knowledge Distillation Papers GitHub](https://github.com/lhyfst/knowledge-distillation-papers)
34. [Geoffrey Hinton Papers](http://www.cs.toronto.edu/~hinton/papers.html)
35. [Does Knowledge Distillation Really Work? - NeurIPS](https://proceedings.neurips.cc/paper_files/paper/2021/file/376c6b9ff3bedbbea56751a84fffc10c-Paper.pdf)
36. [Llama 3 Evaluation Details](https://github.com/meta-llama/llama3/blob/main/eval_details.md)
37. [GSM8K benchmark results: GPT-OSS vs llama 3](https://www.byteplus.com/en/topic/577637)
38. [Azure Llama 3.1 benchmarks](https://www.reddit.com/r/LocalLLaMA/comments/1e9hg7g/azure_llama_31_benchmarks/)
39. [Benchmarking evaluation for Llama 3](https://generativeai.pub/benchmarking-evaluation-for-llama3-55d9fcf10dd1)
40. [The Llama 3 Herd of Models](https://ar5iv.labs.arxiv.org/html/2407.21783)
41. [AI Knowledge Distillation - International IP Law](https://jlrjs.com/wp-content/uploads/2025/08/86.-Shristi-Sharma-1.pdf)
42. [AI Knowledge Distillation IP Issues](https://www.pinepat.com/en/insights/columns/ai-knowledge-distillation-ip-issues)
43. [AI Model Distillation Attacks Explained](https://www.mindstudio.ai/blog/ai-model-distillation-attacks-explained)
44. [Knowledge Distillation and Trade Secret Law](https://scholarship.kentlaw.iit.edu/cgi/viewcontent.cgi?article=4489&context=cklawreview)
45. [Five doctrinal pathways to invalidate the anti-distillation clause](https://medium.com/@gastonrey76/five-doctrinal-pathways-to-invalidate-the-anti-distillation-clause-2fb5ba818eea)
46. [DistilBERT arXiv paper](https://ar5iv.labs.arxiv.org/html/1910.01108)
47. [Introducing DistilBERT, a distilled version of BERT](https://medium.com/huggingface/smaller-faster-cheaper-lighter-introducing-dilbert-a-distilled-version-of-bert-8cf3380435b5)
48. [Understanding DistilBERT in depth](https://arunm8489.medium.com/understanding-distil-bert-in-depth-5f2ca92cf1ed)
49. [CSE5539 DistilBERT slides](https://ysu1989.github.io/courses/au20/cse5539/DistilBERT.pdf)
50. [DistilBERT: Distilled Version of BERT](https://zilliz.com/learn/distilbert-distilled-version-of-bert)
51. [Hacker News - Anthropic ToS restrictions](https://news.ycombinator.com/item?id=46578701)
52. [RepoBird Terms of Service](https://repobird.ai/terms)
53. [Hacker News - Anthropic ToS limitations](https://news.ycombinator.com/item?id=46579226)
54. [Screened Terms of Service](https://screened.io/terms)
55. [Artificial Intelligence's Fair Use Crisis](https://digitalcommons.law.scu.edu/cgi/viewcontent.cgi?article=2980&context=lawreview)
56. [TinyLlama Pre-training](https://kaitchup.substack.com/p/tinyllama-pre-training-a-small-llama)
57. [TinyLlama: Open-source compact language model](https://medium.com/aimonks/tinyllama-open-source-compact-language-model-rising-from-llama-2-5238cbc6a0e2)
58. [Llama 2 vs TinyLlama Comparison](https://sourceforge.net/software/compare/Llama-2-vs-TinyLlama/)
59. [Llama vs TinyLlama - Slashdot](https://slashdot.org/software/comparison/LLaMA-vs-TinyLlama/)
60. [Aya vs Llama 2 vs TinyLlama](https://slashdot.org/software/comparison/Aya-vs-Llama-2-vs-TinyLlama/)
61. [OpenAI Row Terms of Use](https://openai.com/policies/row-terms-of-use/)
62. [Knowledge Distillation Review - NCBI](https://pmc.ncbi.nlm.nih.gov/articles/PMC8053015/)
63. [Legality of distillation - Law.asia](https://law.asia/openai-deepseek-ai-distillation/)
64. [DistilBERT Size Reduction Data](https://www.analyticsvidhya.com/blog/2022/11/introduction-to-distilbert-in-student-model/)
65. [Google Gemini API Terms](https://ai.google.dev/gemini-api/terms)
66. [AI Terms of Service Fine Print](https://www.termsfeed.com/blog/ai-terms-service-fine-print/)
67. [Teradata AI Features Terms](https://www.teradata.com/legal/service-specific-terms-for-artificial-intelligence-features)
68. [Google AI Principles Update](https://ai.google/static/documents/ai-principles-2023-progress-update.pdf)
69. [AI Model Terms of Use Enforceability](https://www.repository.law.indiana.edu/cgi/viewcontent.cgi?article=11576&context=ilj)
70. [Unexpected Externalities of Distillation](https://antidistillation.com/blog/unexpected-externalities-of-distillation/)
71. [LLM Distillation Attacks: The New AI Extraction Economy](https://medium.com/@adnanmasood/llm-distillation-attacks-the-new-ai-extraction-economy-20672360b586)
72. [Distillation Attacks on LLMs](https://gpt.gekko.de/distillation-attacks-llms-security/)
73. [Adversarial Attacks on LLMs](https://www.computer.org/csdl/magazine/sp/2026/01/11369832/2dHi7f004xi)
74. [Threat Intelligence: Distillation Attacks](https://cloud.google.com/blog/topics/threat-intelligence/distillation-experimentation-integration-ai-adversarial-use)
75. [DeepSeek R1 Distill Qwen 1.5B vs Llama 3.1 8B](https://llm-stats.com/models/compare/deepseek-r1-distill-qwen-1.5b-vs-llama-3.1-8b-instruct)
76. [DeepSeek R1 Distill Llama 8B vs Llama 3.1 8B](https://llm-stats.com/models/compare/deepseek-r1-distill-llama-8b-vs-llama-3.1-8b-instruct)
77. [Exploring the potential of DeepSeek R1 Distill](https://medium.com/@bonnyjames0830/exploring-the-true-potential-of-deepseek-r1-distill-vs-e326b1995073)
78. [DeepSeek R1 Distill Qwen 1.5B vs Llama 3.3 70B](https://llmbase.ai/compare/deepseek-r1-distill-qwen-1-5b,llama-3-3-instruct-70b/)
79. [Benchmarking Distilled Language Models](https://arxiv.org/pdf/2602.20164)
80. [Llama 3 Model Comparison](https://www.scribd.com/document/877562284/Llama3-Model-Comparison-1)
81. [Llama 3.1 405B vs 70B vs 8B Quick Comparison](https://myscale.com/blog/llama-3-1-405b-70b-8b-quick-comparison/)
82. [Explainer on Llama 3's Performance Benchmark Values](https://medium.com/@ingridwickstevens/more-llm-acronyms-an-explainer-on-llama-3s-performance-benchmark-values-36722c6dcabb)
83. [Meta Llama 3 Launch Part 1](https://synthedia.substack.com/p/meta-llama-3-launch-part-1-8b-and)
84. [DeepSeek R1 Distill Llama 8B vs Llama 3.3 70B](https://llm-stats.com/models/compare/deepseek-r1-distill-llama-8b-vs-llama-3.3-70b-instruct)
85. [Identifier-Organizer-Adapter framework](https://arxiv.org/html/2602.12172v1)
86. [Evaluating Knowledge Distillation Loss](https://arxiv.org/html/2604.25110v2)
87. [NAACL Tutorial on KD](https://aclanthology.org/2025.naacl-tutorial.4.pdf)
88. [CollapseBench: Model Collapse Evaluation](https://openreview.net/pdf?id=c1yJSRahCt)
89. [LLM-Synthetic-Data Papers GitHub](https://github.com/pengr/LLM-Synthetic-Data)
90. [Benchmarks: Llama-3-Distill vs Teachers](https://www.distillabs.ai/blog/we-benchmarked-12-small-language-models-across-8-tasks-to-find-the-best-base-model-for-fine-tuning/)
91. [DeepSeek-R1 Paper Analysis](https://arxiv.org/html/2501.12948v1)
92. [Specific benchmark results for distilled LLMs](https://aclanthology.org/2025.findings-emnlp.451.pdf)
93. [MiniLLM vs Llama benchmarks](https://llm-stats.com/models/compare/gpt-4.1-mini-2025-04-14-vs-llama-3.3-70b-instruct)
94. [Small LLM Performance Benchmark](https://ascentcore.com/2026/04/01/small-llm-performance-benchmark/)
95. [Mac Mini LLM Performance](https://www.popularai.org/p/mac-mini-llm-performance-in-2026)
96. [Benchmarked 12 SLMs - Tunability](https://www.distillabs.ai/blog/we-benchmarked-12-small-language-models-across-8-tasks-to-find-the-best-base-model-for-fine-tuning/)
97. [SLMs vs LLMs in RE Tasks](https://arxiv.org/html/2510.21443v1)
98. [TinyBERT Distillation](https://medium.com/data-science/large-language-models-tinybert-distilling-bert-for-nlp-1a928ba3082b)
99. [Reddit - TinyBERT Discussion](https://www.reddit.com/r/MachineLearning/comments/daom3a/tinybert_distilling_bert_for_natural_language/)
100. [TinyBERT Findings](https://aclanthology.org/2020.findings-emnlp.372.pdf)
101. [TinyBERT: Distilling BERT](https://arxiv.org/abs/1909.10351)
102. [Comparing Top Small Language Models](https://www.c-sharpcorner.com/article/distilbert-albert-and-beyond-comparing-top-small-language-models/)
103. [DistilRoBERTa Overview](https://www.emergentmind.com/topics/distilroberta-model)
104. [ALBERT, RoBERTa, and DistilBERT](https://towardsdatascience.com/everything-you-need-to-know-about-albert-roberta-and-distilbert-11a74334b2da/)
105. [Emotion Recognition with Transformers](https://arxiv.org/pdf/2104.02041)
106. [DistilBERT Architecture](https://zilliz.com/learn/distilbert-distilled-version-of-bert)
107. [Comparing BERT, RoBERTa, and DistilBERT](https://medium.com/@cd_24/comparing-bert-roberta-and-distilbert-for-named-entity-recognition-in-google-colab-1a2b6feff953)

**Sources:**
1. [lau.edu.lb](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFSICF7Dbl2cI6w5zMU2P51srfbkOt1akZkg_CTVGT7X1WxiIF5KcTfC1a_lIRQ7fEN8iL4UxTYudMdsofYVoudjrTedDQQs8MUeZGZtN_cDKDy2pL88aB02cMBXE0J7npOqCC0QMU6Y6nx_u8fOL5jxlB-N4O0mn2y0wDmVFb8)
2. [uplatz.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGO2vaBqTfGG8J9o-R0RX9uY014Xc83910XuN__mtdzBAchYoiPlns1FRyRui0DMkyU2hE8RrZ50BQgqbY2C-hvud295sD4FR70eQitWfJuTr1KcLaxSqEibq3bvDApvR7mPmPd_r0B_jYq5YfsvilMgdzwnbLQiRZzh_iLygFtXyXFSxXBDXsdmL8xP8eZ2Z5q3eojzMOi5UWk_OUXNaQat0X_ysJsnrCyNXUmAZdZwaMjl5jpll79Jf2i-I4TvE5cBVlpkHHHSJskOTaVtZoPygtx6YoRXR01Jg==)
3. [aimind.so](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFukNdZdA0w-M8QfM_9Iz3EM_3uqZndSbfJ8Ss8poSRuSwTU-K0I7-E0siXVxaIcvekfLI-l0RxNa148gknGHm77GqG4s8cxuptbh0EhYmK5PMzYCpNSixQbU7i-KO5DMcuVliC-bEi2GLvA5ES9n7U0nYYkvz6b5vysz2eSMu5pmeE8DOrVozM1Ca_t1nNLJzyxd7kC1sXx7CdEBIczVgFSq1pNWmPsaoUj5w=)
4. [nih.gov](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGRRS9-UvWe4iMZ4_BbDLgkJhsrjY_2de8gjOlVZFNREG5X_wtukz5ZMhYlpEU5wMEKlGxmTxnuLbGc387PrlT7tdJtEmhFm1xCo9dZEWgErfOHKKB-GyknnrooAapuM0nvrfY6I1y8)
5. [escholarship.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHzvoVK3OgEudMef7e_8xSQbGK8QbooOZchpr__OgVgngiHbKNkBL2a-ThLY-heTuqQkOryjfvt6VvNb07EsuNbyr4JjMGaXWmRlkpIVERkcnODVm5UVtSEg9NzPnbXoh9RhTka1WBtM5E5WUnE0hwU78LVUp8v5chK9K9608irNVK8jC-t3odDQtpeuJ8uF4CpjbJvhjpN0WDrsK23c9f0c2F5)
6. [ed.ac.uk](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE46bkZ4ECBJMmYzeEQPKZkcKpYyxEfW9UI-fLAs3Lv4ExwzPpPnGZUnLvFzHLxltqpDmTZKw3VnnNA9Y0vqjywySIL6cW5VegaqLc5MG6vSDDxXBxu0c5hoJqbIb_dWYlH9H0ApIYLv-jV4hBIGzt5cTHvoA==)
7. [zilliz.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGpQiBJ-QsrWuWI0IKQ8_UB62_oiFDin4vZXkBfA54fWtUCR8KSK9tXp-ocz3xP2XLF8OTTt5m8dP3RaQVCzLScELXbGiGnk_RJpx-w0XhZZVfkFHNsBsAFZyv6fq0wlzdH1iuPe7ZwtlO6Ma7TjB7fPH4i)
8. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEWI2FtvU-_djMmu89qdyJqd_R0SfMaUvJeXQ367atk0ETzrDBoTxLrMod-kFW1dgKrsJd9fxCg8ncaUA4C_T3wS6LJJNAKkfOe28iFg2ohSzsn33Qafxy7TKW7AbXTs7sREcT4hl7UVjg0sYpA07nvMmbTo4hjEUGTnd-4HmGQfgwkZ_0C5Twwsg0hljsOacxX9E0ReTTrsiuysOuZRsUFGkwn4C8JOk5NliKu6CNCJxPIZVk4ioGwm2pMrmMsS5vTcW_t9is=)
9. [pinepat.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFhqCJBfjaLi5QDSO9UL9kDKE0ptw3iNhBsDq2hqN3sDKfUYT4rZv7d-DITu47vo8MnmX_yHJlHx1QoDCwAlY6ZTiK2XV-OsBhVBnq6Q7ThK-I4n1rAYp9sjm02HQ9TuwLEPCVGGgBJPL1tdmgm5OdtSXA8LAiENp1gWhj6_gWwKiOmUoUH)
10. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFgl00hf6nlZh489jMCOSotzv6MTIADs_OwOL_tWWt3UqOys3wupk8tEL5CT-1XXBF8LfrAL2WKjpgedzgoZKuAivUbPHfd8GR9Z7Z1uwxlMcObcXKNtx8zMeoQd3bjr98xi7RohveT1A5HFNz8J-WHD3AkWu2PX6wKuUXBlNASijda)
11. [github.io](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEiIWAM8ysioaguBS-XMV9W1KpCYl4Wgwmfa6YxENrDhbiTzjoX15yyK3uQN-OumM3BgojihTH2xLcWsz7j535BPYPjP2vmp9rTtFNTJ3VkvthedDyNr-MvnF9e-agg19eqolT3BuK-jHUML6GJrBMy2V0a)
12. [analyticsvidhya.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFzCoFMlnoQKr-v-ARRPQz3BRm0IA39q34yd2TrWdl8Xmwv52gBVUqK8_0UBAcZQFRZIwOWkimB21EvIsbVXhlQ1TjNpVMdLmxdXG42SZ8K5CvRAEKyoU1IRGno_8VdA-WcJAwzY6P84dmXmhu-bsPgm9R4Glf2Y7OAy9pwgCSJTBzY9Uxli9eWsSujKhgxfQ==)
13. [neurips.cc](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG0h3uoEbmVdqaoBvElGoLYhSL9FukqcsjoVk0_T_0LjVXW1bUPrHf9tzypcKupcsxbcckHliuUPPHwhSwsjCtuUIRDi62q0N4JzNSk-u2fAbrWMG_ZRdhR72AJiC6o2_uBPsiVSiUR-jZWGKdJWkCRRlG1axKRit5qZLnxCBqzJXl-Js1ENGtpMzGIqSiY82iKGESsfiXolf6FXQ==)
14. [emergentmind.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFdDp5Y4zkiCiXJoMpmcgKNBOyWNHNjubwk1aBBvu7GaBgqnsP4XampmtpiexXhr1rUFoyiqyrB-RUNNFOcUcVP9_OrVz9WrKRFYFqkSENgLeSfIaMifoeLmbWqanso9kd_kJfDdVbsERZKZGir)
15. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGwgecOgy_5m78094vCFuIrqy4wX2A055Q8Z2l2wYy6P3XrmPXb0rG6LdyLV1mVvv81yz-ZOQpsN8s8-LPSZCQd6ckJpS9Vd3Obl3hTjxuIv2JbGKLHjL9RLI6S_UHBpsX-IAKH9BZS7Gt5PWbFYB8R65HFOvXdaL6KrnrsNKrCaUIaODyg5tAxgMvIz-8m7AcCzmUVHl0umfCBkDIeGztPb4-Rj5DCT80WskDqqcPCr9K1P5gToWdntw==)
16. [openvino.ai](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH-ZTqPVVfI1AmUlGe5vZV2aW9J5mQLrAfHFqwntjvj9PC1E8dASIXNjgQg8n-M7xYQwGSMkwINCWv96KK54wDhYh9uhHZlCdcINEImtyxoGN8kDQGUwIKWa3elDMg4iaSa5f5EYw8V6GHmCRm4A3qWaN_IATu4lW3b9u36VS2p8QWF3evkBynTuQPCItLq2YmpTgcXzCKAN5MHa1hQWpluKvMxJqSxcGAd9o9nbg==)
17. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGUQNP4oYMjX2yGfpyREXge98MCc6WUhCNpY2LaOVv2A4X5Q63CRHULVG92E9RjBvxD0m8Uutn_jzPc2GLYVSgQP9fyzQIcxfgKK4Xt73dQryb3FDDcQ7RIUQPJseXRUDgckEZtea3aUuGXJl7iRCqs3fzTr4kVGooeq1aSbZCOWx-HhwdtzUO4k1wMswXX6vLMSHBoxiOPe-GITr8-ZDny8md6_8h7h7S-dOed7g==)
18. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFf3Y7wey8UWa0AwzLyXlTBGyjBijSw5MGkkxxzfgpjnugyoNSMy-xjCSTsCuu87Uugj-xLHlJqXMh_0CeUL2SbNx9iIx0ar2cBIY_uswg5QlwmxcfSH-Zuojr0wFztT8FOwQrn4197Lw==)
19. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHLY8uoMCao0Se2YWLJc5JcwfZeW71rMndIuLil40VmMz0BW56yLJHiWPt9aldG2bBpgcMgdrMDHIYAPJzWhRQhptt6yA6BC9LVzed66fZvipQlqBe_A9UQ3mDV56mZ9oeK0ypf3e2TN8Y1)
20. [icml.cc](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHFC9C1gQw-TpAyIdWj7QzQ7QJICkrhAvwUMkigHPwCo6PZcaU9JQaSUVikXusH9gbkLMuPGtWizgqZV4fPGjTfTC36Fzf9gvXAAEYtAw1dSupZbL_j92A5gp7I2KuPuA==)
21. [aaai.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF2EFoBL39oN6hOpBX9zvAzmOtOAHrGCbXnypsYl5CyDKKkMhlGU-mApXXI-ucqB4G_nKaqPFI65QJl0T1LNWc-KnO57dH1ZzlWOWhBS54imthuf4Hv0cOqdigLooxHYGd1690dPhmHSPtrbEBZDK56rjo=)
22. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEuHjKKkFElGVxblnDt9mcgzHBMR602kOyDIBw4XTE9tLc0keN1Asa40-Pyi7gqogVtAsXPaK8QZaU23rWzTF0PU60q6XZ-4guLVHs73t9O77Qqh3i28d4-Ww==)
23. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFyAETB0mVsBetWFLO2Va4MZ21G50TtnHhmz9aG3o0h2EIAduWOfKV3kNXhjYJK_3XfGSeFQGF2eFBaFuIBwC4W3aSiyLGtUxpHCJp2DhS63IDR8y_mcNoMOIyUrghsqa0=)
24. [neurips.cc](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE8KbgCrmeCsiTU8Hv270lTgK3RDQ0YqWPTI4JGR9vlqXcJiDQYaEY6whMNg-_wGSTAxTbbwT0qbZaSoKA91mzn8-LSFnhjAdAEErd4mbJAOiUDhxUE2t-BvA3TAQ==)
25. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHc5D2Dm9XQFK2e7m_z1HPWPD_dqOPAo79H8K5M0cVWkg9nYA4J9Teto4DMjqFR0AD1JQjvX9njmme7mjujxdASIfRfRsf1w9w1d-eRF-Gsgsal8ohZgTS77QZbF-g3vlc=)
26. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGD4GAY96eeKUz-Syq-oC1E51fMIXZGNruKx2ZYPOezWSu8ikzAw-gRYeTfQ_t8Gvf5ZSMTNBJqpu5-dUO-ysYT1ZwhK-j_nuOD9rJG_nBfVq1E12efHA==)
27. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFfICxHCSzW9C7MWwQUNFjrmMD_oO0q0fC6f-rUDBnRdn2uaVmRExz-AfkUydPN8ll74g3-IHZw2_0ymdeMIrTsJqIqN-OCWrCqtkul3KM6skCtWxdvv6gwY33b-j5FZ6mdgDLg3gGMhRfrf0dDef7DUPUamWg1wQXa8iE5p6LPyDrUyS6QosG77uf9nhi9FeOsYd9fax7VDuLuyEocpyPeRH2uSj-ZdxsydpVPlybeJg==)
28. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG9mXRbcVu5RHquAJyvMrorgwUgIYiIgezZT5pTZyo0rBAmMnPh_IJlYrZDJEJpBv579WRT1IC8AqBZ-1QDgj72ipdgkS4w0Ox3TQ-OAgC4Z2wyCLysOKcXJDLd9oCgqofEpQ==)
29. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFsT9SbNRD-f9iW_-YYUAH7FGWm-ypYZv6Vd4YQHCA2OCDgbjahV6puAnOA19-7KiEbQ84cAU6hYlA3-hYB1sqAucuYtgiMXrH3_CfFxRJHF64YKWBqA_EQS1DkclXrHDKLIXpJKnSfAgAmgyCeZBBifbsbPWvxmkBu_xtnqbahLUuh92Y4MbxGjhE-Gtr5ZSQs7AYd5sHUmcIUQ3rXHIvxD8rlTO77obyxp3JpLV2M)
30. [towardsdatascience.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEyt6_m2ryTvYWTu8tIfKdNZL44UJNuCvB1Myke-iQNsqbNcjWLGS9ipb_bWQZ8RL5DoPoWDTAOe9sri4jb1e3-9TXyj08G-w6uYvhsrCLr6yx0vfMy1LaG-zjc-fVagFxKps8GMg1-gIymnKQHOThE01SjlMCGRvG0bH6L8Pzi8oXPXq7E3tYkyftwwyOVGuBJBo5yyDcNhJh0yjjteJ_vn8I=)
31. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFAd_vcrhmSrXwCRKgTYGg09uUCM9PDRGGldB6IUL2bKFS-PHMXwtL7x83pSgn5N4TJptrq9UTTR8KLBBmtdHze3q05vG7YGqno6JAFL4qDZKFYuy6vP-nwqSLeMTWdWGuBdCB4--6ssbJCXO8q2vo-bSggA1O0hGz0QYYN8M3ZEqMNAGP4RQTX--uaS7h-1TDhBWUYMI5TRMA=)
32. [reddit.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHKkmHYvopbFo0HQrsC86wgeChzgJ3If-5z1rvRnrnf21hz0LjCxyHY-2Nt6HPfqrvzGSan3HO6nENc3tq_go8lLP-Sc_YLJhVuigkYVPDPhFepCq0iKe9LbL1dejzsNgkgoIYbeqtnGhRTbdYrmWKWtFygveOvJnXyFAcEOfXVK7tQFOe0HAvWXFPiaIr2F-3oUdGKx-4EzPDmIPYH)
33. [aclanthology.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEmH8LCHPn837L6NcLx8FxPIF-C3uIDOLhMsWZwck8HAev1CeUIAkI8DYuQwecKtXPbkySLWrrib3OQMlTlh3k334j84qoUiNoAXfR7WXfMERWPInd21Az_YxpgN9e_CYZISHsXuHmWfc3n)
34. [github.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG_V2AXBvmLxgNMvL2EATdZInb4RCRFNG3KRLWf2uyKFbNt6ycFcUX2IpQLWFZWYfFViCvTIwN0E_hCwxfDBv4W7IG4yjJDRKJCbUHxNVi_ljL6t_pjhxg5w05iRPNNI6zlpZLr2Y3bFyoe669_LkS1m_oxuQ==)
35. [myscale.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE8uC1HTCY48vk6TULjbX5hnNNVBB4RCduJXEOPnhB6e2ZqhkzlKNjoTiNiXmVG3WS0hS4wJ1Esb4H2DTqprlIBff-43Pi62mVrxHct6k6NZlEXfuEWOrevc-IKnvBH_K8DsxKeKY_CxRnhUowqSrYfavPNhW95)
36. [byteplus.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE_iNcMz_kEEkdTZJ4NowT1G-j3CGY8uu49K6iYzqGRfVIsprznbtV8gJX0ANAMQTdqLi80JtDxkF1pU8fn4dogsfUlCJAIO1cIvc1z1ZYqhHY6n1KnxmyFftxA_WdZ)
37. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGQUtGAw9CdOlOGaVK03tQPTEs4LdG_U0gkP1wjtol8EmAF8TQFt7_J_YGd7x8lC90WeNd6JpL0p8XmX4ZbRQ3h1HXmg9uxGetE7sT0wls6Db730ZXKcnSr686TzUDk5nd1Sg==)
38. [reddit.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHAr3h3u7Ub9enQ0n3RxaKBFp0uJU9qMgb6pNSKt-7dyK-7Dl7RtFm0etFBbNHdsZzeZAgnRGC9KK9NvEwIUKkAv84u5A-W9XKSt4JLfX--M9YY4d7uCi24hAqRVPK9Zl-pFilNfLApeDQ80-qrwrX6IzbdeDIMivChxnBTqR6gLHduX3Tz)
39. [scribd.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGx3TjiXZdKYBC0_G2YW7jfKJ_VMR8u2-Jflehx2CZQIoAbIqu9XEsEE0ZmHIERjya9a4LLgtUord0lndyI0dYyTnq2tK3x6H7ZjA6rk70HvARFQ2iZ9iXYLELpqmSNhyXVQjtuN_CdM5vlBmNhj7JPqDd2LMlNTVwA)
40. [substack.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFktsbN_Pu4sJjipeYJzRXRaodPyVELoc9eUApC6yORhOKmXJjb1JCF4Q2C5miPzrbH0iksJ17R8Ae2MwPdJDyOgKnB5ON2bw5LvgffZtdIIlRZptGQVZwF7-9Q1He_O05lXjlacTL88dpV5sWSBMU8hdAIcRrwQwyGuw==)
41. [slashdot.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFrj7wfOQffGu03yoG5B3aIMOmPQTsj87ztCIMaNaSeI7zSbyntgnD_Z0tLDcjV9fc_1xlsNDd2MswSlIh8SzYQIlYznLkxjqgIYvMXE_n1Y97H-CXWbjf3GXwnAm-QxO12xrSPs7UPRGohRbsz4mymInM=)
42. [sourceforge.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFzfBz_7P1QUivxpjb1d5hTnUUzoW0bKUjJlw_Gi-UWQQg77MRr53-6dUnTTPHaJTxAyh4eIwtzkVm5MW2ocFjVJeVhpqrGvEWNaA5khvz0_EAlTnQ7UEIabXoGpmGWVdPqAzSGEe7Q7gqIsW3ku4AO3Tr1OQ==)
43. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH_azmyNou7TVX47JCtnfzAvsYt2PtckhxY57gPs_l2zejvnGPpg9x717usVLHfnuMzNnB3dJ9A_sg8Nwwqwff4xMvAjmVFBMU3Tez_loJvvXM9le5KiIeVjeQvswgNCWMw64FpOFgVSKDN8HS3AOJ5Ll1Ap47VvscXyT0dC4zDJHrDB_FkyBB9-bIZWLvrpb1nssDBqnirkenesHWlhQ==)
44. [popularai.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE-RjsQKxf0D7hjvrEAZzS9PbHp6NtwWK2593wyEG3jZicFVoQjVLOEI3ZfoOWRGEiKO3KdTbmEpQrlK2qdngvYvC1h0g-OnFzlyR796I00QVIhikijJM6w-LLNO5rv-UsaEC3DM-9PNsvrUg-ZBF3x1r0=)
45. [ascentcore.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEgHXHpz1-zJQ2SKs4RlzaTmwnh2KMyKSSWYmAfcK1K9nBfBkigBlygaSCY5f-BwrY8iBmQ3BGaGwfyy6VDbatpCQ1ZcksC3HUvdKSBHHczFvU_zVo7_fkmc9dgvAoMki6yWBQpIzypMLcHs9L8mqXzq4Yb6dp6IKc=)
46. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHQeyxcVc7NlszpS9hu4I5CQ9mp2XU9XEyfO01LIm5tMUoNrQMi7aF7VT4gXGFGGPMH3XigIyWz4z1Vzw4oUToGLzz5qkMTg3gvtWhHTdEDBzcJrbIAd-fePg==)
47. [Link](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEHAE9uBk_qu58DPF0kPUGFv9WYgvOSuct-TuU16EMK75zYy1-W_ctR6JHsk_6-WNXwWbKozapG8RYmDwcnGeFcSH2neCrTqro7OnuSPyZZs3n4bHrSv2v-zw==)
48. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHw-0AT0ruTmSH0t_lxJ_HWCeQSxVLhqF5zSkUOVQn9d14hGnkXMOxNBu2EvndDyN2Xz2dxUzQVonjUXjyMBfAGoqZAOTu5MueEqOifoMjSxon4Ko-c_A==)
49. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFbl2k4HyuqRD4bT4XQz8eMTA8eqQlspGPkx5PycBkPgFaMN0nlPWL5YIwGyZEW7NXsllAQ3HiLqUpVNfiKjykQlYsZN6MA_M3j-FWvx6CIdWlJmJlO6Rri0g==)
50. [arxiv.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHUY4mCbv5XFayPBzJt-WAOsDEg9HUqH4lvj5TL3FjpL0Vq0J0bPkatm-NMoo-X_EKHpKHuqy_qxQYWywY5kFegmgE-p6omH9BoupkbUXSuX9D5Aq6YBsiaQw==)
51. [openreview.net](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQED9uaunlAMKjSfBjDZ9n8FbpUmpNtTdFXUMo-UWL_aI_y2NVfXLI6u9aoBckO2jsFRCJtdyvW-8D4b7-aaBdtdANx_SFZcIojah3lHwev4uLXhhlySPH09zwfWLWNQ)
52. [github.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEh4P0TcPPBe0drUsbgJlgud_O1eggz-1ZNleBzESZc5QgCDWsCIKTqlaFAYniYPB7GgomjICQoskg-ZqqVdNzkH_BqXXx52WZYJRDosiXNs4t41KeUxvSyH0OdCQDTBd3O)
53. [antidistillation.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGIhkgzh7yw5zPD7g5quSps-SEUmFU1hMuKyoe67uLk7JYLgM10orBanODtlt9wPZMbS80nhyuVkCA9J-dxZSt0lkH1wO6nz_zbdMOsfsttGlnOpGEeYKQyRPXReoXevwhAq3Mec6_Zl8BuMV6-xbsewa2SouENk-ranqRBHhl5slY=)
54. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFLfEv6evefnIfpRjTPRXqgS9QXRO5Ch0xK2wQe1YZmXtzdrnhBds6TK8__WDFu4W-98WKy-Yp2Xq5N8OKncGDSEKS7n0atG7AobfCuMtbu3lCaEFp280flPLQ9729-D20KR2gJkmgKLbdcaAOPSHbVVzhhAU0VlgR8BVC_1vIMUgqyW_nPhT_zRiE9AV6T3NMQwvpW2a5d4CE=)
55. [google.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFnkMw25PfWtP3SL7ely4BZWePuxDhKHLLy9h8_I1RN9iTv8rljtYdtY2Dz1D8ciFwxnfu0pAczCQLulAKNbC6QaS9H7Qkr_XG0aUAh7KHx4ulwdSxyfGjV5NR3GYi_rFuQYyaDBej33-_dNwwOFQvCD82eX8syYKCSGJ0ehTIwH0r1J0QTj5jA9mduKUm6tYRYmYOWxL1G3l_boWkir6x2MLNPjuIOG15Gxg==)
56. [gekko.de](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHgsDZBDOihpFT4ykUN_yaqMyl-BW6yATSyVaSYMCZmbG9RfyBQvJbpL62Q1CUzJTXW3bx8gGwszhGeogYnvlW2zHnsUGyjyM-Sov7S9oYVSR0j1hMyO3UVfh4GUW_pAQdCanEhdtktYzwdk7QV2Q==)
57. [asiatimes.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFj_0Qd6HT8yrA2oyx9cmKUy_O5M0TA9dKXDJ28-LC1zMjhcT1XgxjrhOTcy7Wj0czJ1v656lQcUs3oUpQ8Bf8vXEc5UFNjCSXrgoxV-nSB0DlhbdZrXKZs67LGS4kWraqjxhU1BiGWOOR08lEtEFCT4YTwNV9kR1WYkY2baxw_1espyQWbPzJRnKcmAbtRQArv434C)
58. [globaltimes.cn](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGwelKziXFMPws83-PMTcFnDPlDht4gBTUaDNd5wzYWQiQAgoWXTjeEKw-shT8HhPMNyO2PK_rg9_os28KTZKzpOlroEmfBWGMICVBJIer24j5_POp-MBywUAmb8YjUjZMB3kI0YnxDZef1)
59. [computer.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF4HwnJaiqcfd7Kkm7iJO4kpXZh0w-9bGpQ6p3cbojRuRha_Uwb98TYLRUKEmJ426BS_Tae7FbcLSAfd7c8dgCsE735x5nwc4rKOS5EHe8iiUUGAtm-6-6qVxH0kyBXgB15prXuhkTrvslgiexO2pdEYjIItpb8TwdikX_h)
60. [claude.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGSlGJ_4T6pRvezN-hc5z_nr2MsEIu54FznStvvHyA-py9uX7JQ0wLCU3WVsmhD2DoCDv46e_Abb50_7H9OKas7fVLi7iX_uB-AIAYlPP65lQFosfdVPMhqf5o2a0oaQI45jtVh8rfB165Pv6ponZpzAFPkSacsfvEGjohmi4q5CNWc4-zbbTAmHe7tWj2CSA==)
61. [fdd.org](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQH-5pJteIKbcZuU6u3DF_q8fKIboglYZBte26kOUM6uslkrkaX8AA3OZjtezvyFN2pdgfMpDU7w5a9pnhtYaNNVHA9lQmjrIgBJS8TBWbBVvAENikuHue1-pjM9qZ0Hs9NDMCm0ZGWdSkzfEZJAYtcUA7Eyx_6U4xKxsrtLzHNU3crU8BBJi4WSWen8NcEvpwzGpcwsCKJiROYDlSCpfjSasuG-OoEO-hxeQiw67FFjcb0AsdzN)
62. [indiatimes.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFL1tfPvPTM5yy0W-YFvzeetehbHG8UoHE31NcPr8BcEe2t3v56dZSGvbKFiHXUID2dllCw9iJI7CBz5fKmApOXyTJduivX46-vr6KRdVeiE-Th02jQ560F-YShcaaiuCcWTrkvSHO7gN4yYtJgNnlanrJfdTP7auQlfzmxesLWGv6-z_gSibw57XEI4JIGl2ST7PQLZG1NuF9eEv-VbG4tw65e233uT6dejRAQ4NDM_tX-XUSCjkLWwRlYbsREdCKMcizKFlplvc3K-cJLAlImW8QxeJ2HG74MgzHrz8ygPX-dgzJ6g_QC)
63. [jlrjs.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEfavu3ugOPacOEE2d8RBaiYnc5Ms8pT81knNySkdB5yCzlPbRfef67bo-VyBfXXlcNiBaKNdhd-RHGDwnPjxVM6Ong7I5g2zn2mAZV0u2bTc2z-gxQezkIbvDlizF-p6sFv7fZwVKDTwMNPaPGWmHF3wGY6CZ_d01nQ2E=)
64. [openai.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFSvDAXQKEGhCKvX3HiWgOH8wwH7eLA9N0e5PSePdbN4Vn3Xoan7ftyrC0FTuB7aUsjdNQOhQEB4BcdSSMwP_YcYw_Dvt39_nJFvIGaVAg-by8AZ_juMYthlstuaMNF3d66u2E=)
65. [ycombinator.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGWIvyp-_ge3HruUQGVd8LaI40j7xEDWbvO3smO8d3mSPTOpmDyJGU4KOEtPJCVGsfTwJzOBkzoGii3rxwjX9ZteUtmFMu2eD-zD-8kNaeh5eLe8Ue5gls39lRltGqNQg0avqI=)
66. [ycombinator.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG-kgOwQmMyXnx7amQeQb7L38gdWPP0JLVweAz6ptwA1G16nds4HOgOydWyUoVHNSX6xY0JjXm3zf39-R5tlu12yKxKWhc8SjWJvtpaONA8Vluya5ZJxp8cg5ePAXSx1r0q7Dc=)
67. [google.dev](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQE-cqOhUat3hETW1Xo9pxycooiIIKxqcGl08HU0SMoMsYZco8XbYqzQJaqWalNP10JSjOLb5jxnL2cwM_UHqSQT5x5MYl8nQPgZmrULyESdFrXYDWVgMaFLZ1e7QA==)
68. [termsfeed.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF3dGU0P464furC1UWUF9xrGQaIsWzuT30FJEWwUwem8WaPmfPZFfWzIKMCT5rD29S3WDwGoqTc0vmf5i4-XGv_uUedhWH_mPkZfG6ejQUTrhDW0luRE8_gqQK0YLqiRy8PH1NKFTV8o8H_gbWqJPYiTg==)
69. [mindstudio.ai](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQFdHYDyU_ng1fwDiDVMNSmGb0-R-AEGqA-4HmfMTZRKSmteUHbEZTPyzogiLwRt_g9LCvdwKq7MmTpkZ16JUl1WCN8RGcYT25TR0yRoUTns-TgCD7DmNPoEAxb9mmOD3Uwb3AdWvIEzpYCVrox6HvC8EDz8dAtPMtZjiXsb)
70. [law.asia](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG5kAFrVYsup8j_f6HxXyFAD9TsQFgs6nW6K3mjWkcsFv7977SX3yrKznt_YmLQzraSMvb2sSWO5nGGKRDF__YkZS1ymsTKg9bHbMS4YrAzUDhITA2YLDQeQtfdPuCmQlrLtXPA7np6)
71. [indiana.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQG1BVAjP7ODl_Szgc5-hhZHKw7Q5FsfG9-Pq3dLLATLSttgLkqqQcHNYIE6naFtQrhwkrWnohLXJVPqLRhp8vkHl3agm9HiOIB_HtusfX0gc6z57xcKVJxJBxH87TPkIc8BNrUWqtSSdWCOjp97nWgBRB63mzxY-85q_EP9YdklI_Ymt1rqUyziRs4=)
72. [iit.edu](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHaxVjkrY-KV-kY7d3lz1fQUz8y3p-fj4TRqdpqvOa9ExbntYmIMxqbqNa1c9pke7dBxlw_KGsyw1dCAh9noTP042VVbNraLAxuRTyWVKF-JCl8R75Gtd4kvgtZR2nQ571Kk13OEHGG3JGK0_nG4udbvSFThOu8-K99lEQuVQJcDaF57SvJE9xue1aoGrf-)
73. [medium.com](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEFc08gw00jY-XRhMRh1UX-wXWhs_sRYKmQ_u_AlWpdmO_D_Cd3EmxBVsi224vHU80CiPM413vjlLBj6vdAleHEbm_QQ5YVyDeBQMBQUl5TRBLmxQJl4GBTjd1n7zuH50A4VidJYybiEGIwnsp6r4dt2T5-2B2K0kSvE_dDxlhg3M5aAXaWPkhW5b4UirRBvD-fV7cBv3pPg6AWYTj4nOl09_Ikk5k=)