Reading

Multi-vector embedding models, such as ColBERT, provide each query token with its own independent opportunity to find a match within a document. This makes them highly expressive for search, but the cost is that each document generates hundreds of vectors. A corpus containing millions of documents may hold hundreds of millions of vectors; even with 4-bit residual quantization via Fast-PLAID, 200 vectors of 128 dimensions require 12.8 KB per document, totaling 12.8 GB across a million documents.

Token pooling compresses representations by merging similar token vectors through Ward hierarchical clustering. For example, with pool_factor=4, a document retains only about one-quarter of its original vectors. The more aggressive the compression, the easier it becomes to lose discriminative information critical for retrieval—a risk because standard training objectives never required the encoder to adapt to such compression.

In a blog post published on November 11, 2026, Sentence Transformers maintainer Tom Aarsen introduced and implemented a pooling-aware training method. Its core concept resembles Matryoshka embeddings: applying retrieval objectives simultaneously across multiple representation granularities. Specifically, each training batch contributes two losses—the retrieval loss for original document vectors and the retrieval loss for pooled document vectors—averaged with equal weights of 0.5.

In implementation, MultiVectorTokenPoolingLoss wraps existing multi-vector retrieval losses (such as CachedMultiVectorMultipleNegativesRankingLoss) and accepts a set of token_poolings parameters. None represents the original vectors, while HierarchicalTokenPooling(pool_factor=4) represents the compressed vectors. During training, Ward clustering uses a straight-through estimator to pass gradients from pooled vectors back to their member tokens without differentiating the cluster assignments themselves, allowing the encoder to learn from retrieval errors made with the compressed representations.

Aarsen trained two models on answerdotai/ModernBERT-base: one using only standard contrastive loss, and another using pooling-aware loss. Both were trained for one epoch on 120,000 MS MARCO triplets with seed 42, matching initialization and optimization settings exactly, differing only in the loss function. In the NanoBEIR@13 evaluation, the standard model retained 94.6% of its unpooled nDCG@10 at pool_factor=4 (49.03 vs. 51.84), while the pooling-aware model retained 97.8% (51.23 vs. 52.39).

Notably, the benefits of pooling-aware training extend beyond the pool_factor=4 used during training. Gains persist under fixed caps (reaching +3.48 percentage points at max_tokens=8 and +5.78 at max_tokens=4) and adaptive distance thresholds, despite training utilizing only full representations and proportional pooling conditions.

The applicability boundaries of this method warrant attention: all figures come from a single self-test by the author, the unpooled baseline itself was relatively low (51.84/52.39), the training scale was limited (one epoch, 120K triplets), and the evaluation set was NanoBEIR@13 rather than the complete BEIR benchmark. Aarsen himself noted that no hyperparameter search was conducted. The method has been integrated into the Sentence Transformers library, with code and model weights (tomaarsen/mve-ModernBERT-base-token-pooling) available for reproduction.

Amazon Web Services (AWS) released the Universal Agent Evaluation Framework (UAEF) on November 11, 2026. This open-source system addresses the challenge of inconsistent evaluation standards caused by using different AI agent frameworks such as LangGraph, LangChain, Strands Agents, and Amazon Bedrock AgentCore. Previously, teams had to rebuild metrics and scoring logic for each framework's specific trace format, making it costly to switch tools or run agents on multiple platforms.

The core mechanism of UAEF involves adapters that convert execution records from various frameworks into a unified 'AgentTrace' canonical format. On this basis, the system scores agents using 36 built-in metrics across seven dimensions: tool calling, response quality, responsible AI, performance, multi-turn conversation, multi-agent coordination, and reasoning. All metrics are normalized to a [0, 1] interval, allowing agent versions from different frameworks to be compared on a single scale.

The framework also integrates personally identifiable information (PII) screening and stores results in Amazon DynamoDB and Amazon S3, enabling teams to save experiment data and generate reports. According to AWS, 12 organizations in the telecom, healthcare, and financial services industries have used UAEF to evaluate their AI agents, comparing versions before release to detect performance regressions.

UAEF is open source under the Apache 2.0 license on GitHub, providing a Python library, framework adapters, built-in metrics, and runnable notebooks. For developers building complex multi-agent systems, this tool offers a standardized evaluation infrastructure, helping to more objectively measure the impact of changes to prompts, models, or tools.

US-based AI company Anthropic kicked off its Initial Public Offering (IPO) roadshow on November 10, disclosing key third-quarter financial metrics and compute strategy details on its official page. As of September 30, 2026, Anthropic reported quarterly revenue of $17.0 billion, representing over 12x year-over-year growth. However, the company posted a net loss of $23.3 billion, which includes a $21.9 billion non-cash charge related to changes in the fair value of convertible notes, derivative liabilities, and SAFEs due to increased valuation. The adjusted operating margin stood at 15%, while the actual operating margin was negative 3%.

Regarding compute infrastructure, co-founder Tom Brown outlined the concept of "total effective compute"—defined as physical capacity multiplied by discovered "compute multipliers"—in the roadshow video. He revealed that when Anthropic began working with AMD GPUs this year, it utilized an AI agent called Mythos to handle the porting process. According to Brown, Mythos enabled Claude to run on AMD hardware in just three days with virtually no human intervention, with Claude writing or porting nearly 100% of the kernels. This case serves as concrete evidence validating the feasibility of their multi-chip strategy, which had previously been considered costly due to the need for tailored software stacks for each chip type.

Additionally, Anthropic emphasized that its compute efficiency has improved by more than 4x since Q1 2025. These disclosures aim to demonstrate engineering capabilities and cost control mechanisms in heterogeneous computing environments to investors ahead of the planned Nasdaq listing under the ticker symbol "ANT".

According to a report by technology media outlet The Decoder, an investigation by ChinaTalk found that at least nine US data broker firms are selling approximately $500 million worth of AI training data to Chinese artificial intelligence laboratories. The data includes reinforcement learning (RL) environments—simulated worlds where AI models are trained on specific tasks such as coding, financial analysis, or office work.

The report highlights that Chinese buyers are deliberately purchasing the same products previously sold to frontier labs like OpenAI and Anthropic. By acquiring these pre-built environments, Chinese labs can skip costly in-house development and avoid wasted training runs. This strategy is particularly valuable given the acute memory bottlenecks affecting Chinese chips, allowing them to close the capability gap with less domestic compute.

In addition to purchasing environments, Chinese labs are reportedly using powerful US models like GPT and Claude as "teachers" through distillation, using their outputs as training examples. This two-pronged approach combines external data acquisition with knowledge transfer from US models.

The identified US firms include Mercor, Surge AI, AfterQuery, and Turing. The deals often run through offshore structures, with some procurement offices located in Singapore and San Jose. Because RL environments remain largely unregulated, existing export controls barely apply to this trade.

The authors of the investigation argue that domains with dual-use potential, such as coding and biology, pose particular risks. Coding environments can sharpen offensive cyber capabilities, while biology environments could enhance gene synthesis and bioweapons-relevant applications. They call for the US Bureau of Industry and Security to subject these dual-use environments to existing export controls.

Despite these concerns, the report notes that the relationships are not apocalyptic. Chinese labs still trail their US peers by at least a few months structurally. Furthermore, sensitive environments in areas like biology, chemistry, cybersecurity, and chip design are increasingly being built in-house by American frontier companies, limiting the scope of what is available for resale.