arXiv subjects

Yingrui Li

Publications and source records attributed to Yingrui Li.

8 recordsLinked to original sources

arXiv2026

When Learned Context Planning Fails to Beat Strong Retrieval: A Controlled Study of Planning, Routing, and Reranking for Long-Context QA

Learned context planning selects evidence atoms before an answer model reasons over them. We test whether this learned selection improves long-context multiple-choice QA after strong retrieval, routing, budgeted-selector, and reranking controls. Our primary diagnostic uses all 503 LongBench-v2 MCQ questions with Qwen2.5-7B-Instruct. The planner is SFT-trained on outcome-selected traces from 140 training and 28 development questions; because the 503-question analysis includes those questions, it is partly transductive. At an 18k-character budget, anchored hybrid retrieval reaches 36.18% accuracy and BM25 reaches 35.98%, while the best direct planner-guided method reaches 34.19%. On the untouched 152-question test split, anchored hybrid remains higher (42.11% versus 36.84%). Leakage-safe routers cannot convert a large oracle gap. Under tight budgets, the best planner is ahead by only 0.40 points at 6k and loses at 9k; planner-guided reranking has a +1.79-point estimate at 6k with a paired interval crossing zero and ties the control at 9k. Packing-order and score-flatness analyses did not identify a stable mechanism. Under this setup, learned planning is a weak relevance signal rather than a replacement for strong retrieval.

cs.CL↗

arXiv2026

Tie Handling Is Part of the Evaluation Protocol: An Order-Invariance Audit for Tie-Heavy Recommender Scores

Offline top-k evaluation often ranks one held-out relevant item together with sampled negatives. When several candidates receive exactly the same score, the tie-breaking rule becomes part of the ranking. A common implementation stores the relevant item first and then applies a stable sort, which preserves input order among equal scores; the relevant item therefore wins every tie. We call an evaluator row-order invariant when permuting the input candidates without changing their identities, labels, or scores leaves the final ranking unchanged. We audit this property by holding candidates and scores fixed and changing only the tie-breaking rule. On 30,000 Amazon Beauty & Personal Care rows, NDCG@10 for a rating-weighted attribute-overlap score is 0.85 under input-order tie-breaking. A deterministic hash tie-break based on user and item IDs lowers it to 0.17. The exact expectation under uniform random tie-breaking closely matches the mean over 100 independent hash seeds, while a residualized attribute score with few exact ties is nearly unchanged. MovieLens Tag Genome shows the same pattern for an attribute-overlap score, whereas item popularity is nearly unchanged. We derive expected Hit Rate and NDCG at cutoff k when the relevant item is randomly ordered among candidates with the same score, and we provide a practical reporting checklist. The same issue can occur in sampled or full-catalog evaluation whenever exact ties affect top-k membership or rank.

cs.IR↗

arXiv2026

What Changes When Fact-Verification Scores Improve? Evidence and Answer Accounting Across Trained Verifiers and LLMs

A joint fact-verification score assesses answers and submitted evidence together. When the score improves, how much of the gain remains if the answers are held fixed? On FEVEROUS, strict score is the percentage of claims with a correct answer and a complete annotated evidence group in the submitted evidence. Across four trained DeBERTa checkpoints and 7,890 claims, replacing DCUF evidence with UnifEE evidence raises strict score by 9.61 percentage points, compared with 1.96 percentage points in answer accuracy. The paired 95% interval for the strict-score gain is [8.77, 10.43], conditional on these checkpoints. Replacing only the evidence passed to the scorer accounts for 7.92 or 9.08 percentage points when we retain the answers generated from DCUF or UnifEE evidence, respectively. To examine how this evidence gain depends on evaluation choices, we generate 470,400 responses from two 8B LLMs on FEVER, FEVEROUS, and SciFact under two answer formats and two context budgets. Increasing context from 256 to 2,048 tokens raises the fixed-answer evidence gain on FEVEROUS by 3.84 and 3.10 percentage points for Qwen and Llama, respectively. The effects fall short of the prespecified cross-dataset criterion, while some intervals extend beyond the two-point small-effect bound. Post-hoc analyses quantify changes in answers and submitted evidence, and show when aggregate accuracy and evidence-coverage rates miss the claim-level pattern. The four answer-evidence score combinations reveal changes that endpoint and aggregate metrics leave unresolved.

cs.CL↗

arXiv2025

STARS: Semantic Tokens with Augmented Representations for Recommendation at Scale

Real-world ecommerce recommender systems must deliver relevant items under strict tens-of-milliseconds latency constraints despite challenges such as cold-start products, rapidly shifting user intent, and dynamic context including seasonality, holidays, and promotions. We introduce STARS, a transformer-based sequential recommendation framework built for large-scale, low-latency ecommerce settings. STARS combines several innovations: dual-memory user embeddings that separate long-term preferences from short-term session intent; semantic item tokens that fuse pretrained text embeddings, learnable deltas, and LLM-derived attribute tags, strengthening content-based matching, long-tail coverage, and cold-start performance; context-aware scoring with learned calendar and event offsets; and a latency-conscious two-stage retrieval pipeline that performs offline embedding generation and online maximum inner-product search with filtering, enabling tens-of-milliseconds response times. In offline evaluations on production-scale data, STARS improves Hit@5 by more than 75 percent relative to our existing LambdaMART system. A large-scale A/B test on 6 million visits shows statistically significant lifts, including Total Orders +0.8%, Add-to-Cart on Home +2.0%, and Visits per User +0.5%. These results demonstrate that combining semantic enrichment, multi-intent modeling, and deployment-oriented design can yield state-of-the-art recommendation quality in real-world environments without sacrificing serving efficiency.

cs.IR↗

arXiv2013

SOAPdenovo-Trans: De novo transcriptome assembly with short RNA-Seq reads

Motivation: Transcriptome sequencing has long been the favored method for quickly and inexpensively obtaining the sequences for a large number of genes from an organism with no reference genome. With the rapidly increasing throughputs and decreasing costs of next generation sequencing, RNA-Seq has gained in popularity; but given the typically short reads (e.g. 2 x 90 bp paired ends) of this technol- ogy, de novo assembly to recover complete or full-length transcript sequences remains an algorithmic challenge. Results: We present SOAPdenovo-Trans, a de novo transcriptome assembler designed specifically for RNA-Seq. Its performance was evaluated on transcriptome datasets from rice and mouse. Using the known transcripts from these well-annotated genomes (sequenced a decade ago) as our benchmark, we assessed how SOAPdenovo- Trans and two other popular software handle the practical issues of alternative splicing and variable expression levels. Our conclusion is that SOAPdenovo-Trans provides higher contiguity, lower redundancy, and faster execution. Availability and Implementation: Source code and user manual are at http://sourceforge.net/projects/soapdenovotrans/ Contact: xieyl@genomics.cn or bgi-soap@googlegroups.com

q-bio.GN↗

arXiv2013

Assemblathon 2: evaluating de novo methods of genome assembly in three vertebrate species

Background - The process of generating raw genome sequence data continues to become cheaper, faster, and more accurate. However, assembly of such data into high-quality, finished genome sequences remains challenging. Many genome assembly tools are available, but they differ greatly in terms of their performance (speed, scalability, hardware requirements, acceptance of newer read technologies) and in their final output (composition of assembled sequence). More importantly, it remains largely unclear how to best assess the quality of assembled genome sequences. The Assemblathon competitions are intended to assess current state-of-the-art methods in genome assembly. Results - In Assemblathon 2, we provided a variety of sequence data to be assembled for three vertebrate species (a bird, a fish, and snake). This resulted in a total of 43 submitted assemblies from 21 participating teams. We evaluated these assemblies using a combination of optical map data, Fosmid sequences, and several statistical methods. From over 100 different metrics, we chose ten key measures by which to assess the overall quality of the assemblies. Conclusions - Many current genome assemblers produced useful assemblies, containing a significant representation of their genes, regulatory sequences, and overall genome structure. However, the high degree of variability between the entries suggests that there is still much room for improvement in the field of genome assembly and that approaches which work well in assembling the genome of one species may not necessarily work well for another.

q-bio.GN↗

arXiv2013

YHap: software for probabilistic assignment of Y haplogroups from population re-sequencing data

Y haplogroup analyses are an important component of genealogical reconstruction, population genetic analyses, medical genetics and forensics. These fields are increasingly moving towards use of low-coverage, high throughput sequencing. However, there is as yet no software available for using sequence data to assign Y haplogroup groups probabilistically, such that the posterior probability of assignment fully reflects the information present in the data, and borrows information across all samples sequenced from a population. YHap addresses this problem.

q-bio.PE↗

arXiv2013

SOAP3-dp: Fast, Accurate and Sensitive GPU-based Short Read Aligner

To tackle the exponentially increasing throughput of Next-Generation Sequencing (NGS), most of the existing short-read aligners can be configured to favor speed in trade of accuracy and sensitivity. SOAP3-dp, through leveraging the computational power of both CPU and GPU with optimized algorithms, delivers high speed and sensitivity simultaneously. Compared with widely adopted aligners including BWA, Bowtie2, SeqAlto, GEM and GPU-based aligners including BarraCUDA and CUSHAW, SOAP3-dp is two to tens of times faster, while maintaining the highest sensitivity and lowest false discovery rate (FDR) on Illumina reads with different lengths. Transcending its predecessor SOAP3, which does not allow gapped alignment, SOAP3-dp by default tolerates alignment similarity as low as 60 percent. Real data evaluation using human genome demonstrates SOAP3-dp's power to enable more authentic variants and longer Indels to be discovered. Fosmid sequencing shows a 9.1 percent FDR on newly discovered deletions. SOAP3-dp natively supports BAM file format and provides a scoring scheme same as BWA, which enables it to be integrated into existing analysis pipelines. SOAP3-dp has been deployed on Amazon-EC2, NIH-Biowulf and Tianhe-1A.

q-bio.GN↗