DeepEdu-v1: Efficient and Scalable Agentic LLMs for Vietnamese Education

What it claims

DeepEdu-v1 has two parts. The first, Similarity Chunk Rolling, is a prefill-time sparse-attention scheme. It is benchmarked against TokenSelect, a token-level KV-selection method that the paper treats as the state of the art. The second is an agentic layer that grounds an open model in Vietnamese curriculum content (the SGK textbooks) through a curated "playbook" rather than through fine-tuning. Against TokenSelect, the paper reports 7.7× fewer retrieval calls and about 35% lower TTFT. It says accuracy is preserved, and is improved on structured-retrieval tasks specifically.

How it works

• Similarity Chunk Rolling. TokenSelect-style selective attention normally performs a KV retrieval for each sub-chunk during chunked prefill. This method groups similar query chunks into clusters and issues one retrieval per cluster, spreading the selection cost across the cluster. The authors frame it as targeting an unresolved interaction between selective sparse attention and chunked prefill at very long contexts. Tutoring sessions can approach 1M tokens, where the KV cache outgrows the quantized weight footprint.
• Self-improving layer. "Query-aware context engineering" accumulates verified knowledge from past interactions into a playbook, with no weight updates. It is explicitly designed to avoid two failure modes:
• brevity bias: prompt optimizers compress prompts into generic instructions.
• context collapse: repeated LLM rewrites of a long context erode accumulated knowledge into vague summaries.
• Weight compression is left to existing post-training quantization (AWQ/GPTQ). The paper treats this as necessary but insufficient on consumer GPUs.

Evidence

• vs TokenSelect, long-context retrieval: 7.7× fewer retrieval invocations and about 35% lower TTFT, with accuracy matched or improved.
• vs standard vLLM serving, deployed configuration: nearly 2× faster TTFT.
• Agentic accuracy: 70.0% → 79.5% on "complex tasks". The baseline is described only as "a strong agentic baseline". The largest gains are on financial-reasoning and interactive-agent tracks.

Limits

• Incomplete fetched text. It stops partway through Related Work. No results tables, base model, model size, GPU type, context lengths tested, or code/weights release are visible. Verify these in the full PDF before relying on the numbers.
• Evaluation domain does not match the pitch. The reported agentic gains are on financial-reasoning and interactive-agent benchmarks, not Vietnamese education or SGK-grounded tasks. The visible text shows no measurement of reduced hallucination on Vietnamese curricular content. The authors describe the reduced reliance on dominant-language priors as something the design is "intended" to achieve progressively.
• Unnamed agentic baseline. The baseline for the 70.0 → 79.5 comparison is not named in the visible text.
• Different baselines for the two speedups. The ~35% figure is measured against TokenSelect; the ~2× figure is measured against plain vLLM. They are not comparable.