Ziyang Zeng

Author directory

2026

Deploying text embedding models under resource constraints is challenged by the sequence-length-dependent computational cost of self-attention. However, existing sequence-reduction strategies remain predominantly static and offer limited flexibility across inference budgets. We propose Jasper-Flash, an inference-scalable text embedding framework combining Elastic Token Compression (ETC) with Compression-Adaptive Progressive Distillation (CAPD). ETC uses a length threshold and a user-selected compression ratio to adjust sequence length on the fly, enabling flexible latency–quality trade-offs within a single model. To mitigate the optimization difficulty introduced by lossy sequence reduction, CAPD integrates heterogeneous multi-teacher distillation with a progressive curriculum over fixed and dynamically sampled compression ratios. We instantiate Jasper-Flash as Jasper-Token-Compression-600M, a bilingual text embedding model. Experiments on MTEB and C-MTEB demonstrate competitive performance across compression ratios and substantial reductions in encoding latency, particularly for longer inputs. The code and model are available at https://huggingface.co/infgrad/Jasper-Token-Compression-600M.
In knowledge-intensive creative tasks, Large Language Models (LLMs) often generate outputs that extend beyond established knowledge, making direct verification against current evidence impractical. Unlike factual hallucinations checked against ground truth, such outputs arise naturally in creative generation, where extending beyond current knowledge is often the goal. Yet prior work debates whether hallucination should be suppressed or embraced without empirically analyzing this unverifiable subclass. On the ideation evaluation side, existing work focuses on individual outputs without characterizing the unverifiable space as a whole. To address this gap, we propose a novelty-verifiability characterization that distinguishes Creative Synthesis (Region A) from Groundless Fabrication (Region B), and study it through a conceptual creation task where LLMs synthesize novel scientific concepts. Through 32,400 generations across three technical domains and 1,080 human judgments, we find that Region A is non-negligible (4.7%) and robust, persisting across generation strategies, models, domains, and embedding choices. A retrospective recovery experiment further shows that LLMs can approximate post-cutoff scientific concepts in controlled combinatorial settings. Our findings suggest that the unverifiable space is not uniformly noise but exhibits empirically distinguishable internal structure, providing an empirical basis for more selective hallucination governance.[https://github.com/YuLab1/llm-concept-creation]
Visualized text retrieval, which renders text into images and retrieves them with vision-language models, has been proposed to avoid lossy text parsing and compress long contexts. We ask whether it also bypasses a known flaw of text retrieval: position bias, the tendency to overweight a passage’s beginning. Rendering 1D text into a 2D image may change the bias in two respects: its shape, which position is favored, and its magnitude, how strongly performance varies across positions. We test this by rendering the PosIR benchmark as controlled plain text images and comparing eight visual retrievers with four text retrievers on identical instances. We find that (1) position bias persists, with all visual retrievers exhibiting measurable bias (macro PSI 0.061–0.168); (2) its shape is transformed: seven of eight visual retrievers favor the middle of the page in an inverted-U pattern, contrasting with both the primacy bias of text retrievers and the lost-in-the-middle effect of LLMs with long contexts; (3) its magnitude is sensitive to rendering settings, as taller images amplify PSI and can even reverse positional preference, while font size has a similarly large effect; and (4) under matched backbones, visual retrievers exhibit lower position bias than their text counterparts on longer documents, but this advantage vanishes beyond the pixel budget and at document scale. Visualized text retrieval therefore does not eliminate position bias; it transforms a familiar 1D textual bias into a more complex 2D spatial bias.

2025

This study investigates the position bias in information retrieval, where models tend to overemphasize content at the beginning of passages while neglecting semantically relevant information that appears later. To analyze the extent and impact of position bias, we introduce a new evaluation framework consisting of two position-aware retrieval benchmarks (SQuAD-PosQ, FineWeb-PosQ) and an intuitive diagnostic metric, the Position Sensitivity Index (PSI), for quantifying position bias from a worst-case perspective. We conduct a comprehensive evaluation across the full retrieval pipeline, including BM25, dense embedding models, ColBERT-style late-interaction models, and full-interaction reranker models. Our experiments show that when relevant information appears later in the passage, dense embedding models and ColBERT-style models suffer significant performance degradation (an average drop of 15.6%). In contrast, BM25 and reranker models demonstrate greater robustness to such positional variation. These findings provide practical insights into model sensitivity to the position of relevant information and offer guidance for building more position-robust retrieval systems. Code and data are publicly available at: https://github.com/NovaSearch-Team/position-bias-in-IR.