arXiv subjects

Besnik Fetahu

Publications and source records attributed to Besnik Fetahu.

At least 19 recordsLinked to original sources

arXiv2026

CounterRoute: Self-Routed Reasoning via Hierarchical Counterfactual Credit Assignment

Reasoning-capable language models often produce long chains of thought when direct answers suffice, wasting inference compute. Many dual-mode models leave this choice to users. Automating it is challenging because routing targets evolve with the policy, initial mode preferences destabilize exploration, and sequence-level objectives entangle routing with response learning. We introduce CounterRoute, an online reinforcement-learning framework that jointly learns routing and modeconditioned responses in one shared policy directly from a native dual-mode checkpoint, without method-specific SFT warm-up. Paired current-policy counterfactual rollouts assign cross-mode credit only to the routing token, while within-mode GRPO trains response tokens. A paired-to-self-routed curriculum stabilizes early training with forced rollouts from both modes, then increases self-routed updates to improve autonomous routing. Across nine benchmarks, CounterRoute better balances accuracy and efficiency than heuristic and learned adaptive-routing methods. Relative to always-thinking checkpoints, it improves macro-average accuracy while reducing mean generated tokens by 51% for Qwen3-8B and 41% for Qwen3-14B. On instruction-following and commonsense benchmarks where direct answering is strong, think rates fall as low as 1% while response quality improves. Despite training only on math and instruction following, its routing behavior and response quality generalize to held-out coding, science, knowledge, and commonsense benchmarks.

cs.AI↗

arXiv2026

CHART: A Harness-Rotation Curriculum for Harness-Robust Search Agents

Search agents are usually trained under a single harness. But once an agent is deployed in a real application, its harness is frequently updated (e.g., a rewritten system prompt) to fit production needs. This exposes a fragility of post-trained agents: because a learned behavior is entangled with its training harness, even a harness update that leaves the task unchanged can fail to elicit the behavior. We train a search agent to perform parallel search, a popular strategy for improving both search efficiency and performance. We find that training under a fixed harness makes the behavior harness-local, overfit to that harness's surface form: when the harness changes, the model falls back to serial search. An intuitive fix is harness augmentation, but simply training on more harnesses does not resolve the problem. GRPO learns from the reward gap between parallel and serial rollouts of the same question: a small harness pool saturates that gap early, while a large pool dilutes the per-harness signal too thinly for any harness to consolidate. We therefore propose Curriculum HArness Rotation Training (CHART), a rotating curriculum that lets a search agent gradually consolidate parallel search across harnesses. At each periodic evaluation, CHART "graduates" the harnesses whose expected behavior is learned and replaces them with still-learnable ones, keeping the reward gap alive throughout training. Starting from the same harness pool, CHART makes the model learn parallel search on all harnesses, whereas static augmentation succeeds on at most half of them. The behavior also carries to held-out harnesses: CHART parallelizes on 89% of held-out turns, against at most 5% for the static pools. It further transfers to a new QA task and search environment, improving pass@1 by 5.6pp over the best static pool. Finally, CHART-trained agents benefit more from meta-harness search than baselines.

cs.LG↗

arXiv2026

Customer-Agent: Overcoming Context Limitations in Ultra-Long Shopping Trajectories via Tool-Augmented Agents and RLVR

Understanding customer shopping trajectories is essential for enabling personalized shopping experiences. However, shopping records (i.e., customer's search, clicks, purchases, etc.) often span long time horizons over multiple years, resulting in extremely long trajectories that pose significant challenges for existing large language models (LLMs). Despite the importance of this problem, existing benchmarks are limited to short customer trajectories, while real-world trajectories from large e-commerce platforms are rarely accessible due to data privacy constraints. To address this gap, we introduce ShopTrajQA, a long-context evaluation benchmark constructed from real-world product information and simulated shopping trajectories. The dataset includes variants of up to 32k and 64k tokens, enabling systematic evaluation of model robustness under varying context lengths. Through comprehensive benchmarking of frontier LLMs, we identify critical performance gaps in reasoning over long shopping trajectory data. To address these challenges, we propose a Customer Agent Framework for ultra-long context management. Leveraging a Reinforcement Learning with Verifiable Rewards (RLVR) agentic training paradigm, our approach stores trajectories as external local files and trains the agent to autonomously retrieve and parse them through code-interpreter interactions (e.g., SQL queries), effectively bypassing the fixed in-context window constraints of LLMs. Experimental results demonstrate that our framework achieves strong performance for ShopTrajQA and shows generalization to other complex reasoning tasks.

cs.CL↗

arXiv2026

Alternating Reinforcement Learning with Contextual Rubric Rewards: Beyond the Scalarization Strategy

Reinforcement Learning with Rubric Rewards (RLRR) is a framework that extends conventional reinforcement learning from human feedback (RLHF) and verifiable rewards (RLVR) by replacing scalar preference signals with structured, multi-dimensional, contextual rubric-based evaluations. However, existing approaches in RLRR are limited to linearly compressing vector rewards into a scalar reward with a fixed weightings, which is sensitive to artificial score design and fails to capture correlations among reward dimensions. To overcome the limitations of reward aggregation, this work proposes Alternating Reinforcement Learning with Rubric Rewards (ARL-RR), a framework that eliminates the need for a fixed scalarization by optimizing one semantic rubric meta-class at a time. Theoretically, we show that reward aggregation induces a variance contraction effect, which helps explain the performance gains. We further introduce a lightweight, search-based adaptation procedure that selects the next meta-class dynamically based on task performance, enabling the policy to emphasize critical objectives and thereby improve the model performance. Empirically, our experiments on the HealthBench dataset with experts annotations demonstrate that ARL-RR uniformly outperforms scalarized methods in both model performance and training efficiency across different model scales (1.7B, 4B, 8B, and 14B).

cs.LG↗

arXiv2026

ByteFlow: Language Modeling through Adaptive Byte Compression without a Tokenizer

Modern language models still rely on fixed, pre-defined subword tokenizations. Once a tokenizer is trained, the LM can only operate at this fixed level of granularity, which often leads to brittle and counterintuitive behaviors even in otherwise strong reasoning models. We introduce \textbf{ByteFlow Net}, a new hierarchical architecture that removes tokenizers entirely and instead enables models to learn their own segmentation of raw byte streams into semantically meaningful units. ByteFlow Net performs compression-driven segmentation based on the coding rate of latent representations, yielding adaptive boundaries \emph{while preserving a static computation graph via Top-$K$ selection}. Unlike prior self-tokenizing methods that depend on brittle heuristics with human-designed inductive biases, ByteFlow Net adapts its internal representation granularity to the input itself. Experiments demonstrate that this compression-based chunking strategy yields substantial performance gains, with ByteFlow Net outperforming both BPE-based Transformers and previous byte-level architectures. These results suggest that end-to-end, tokenizer-free modeling is not only feasible but also more effective, opening a path toward more adaptive and information-grounded language models.

cs.CL↗

arXiv2026

Stepwise Penalization for Length-Efficient Chain-of-Thought Reasoning

Large reasoning models improve with more test-time computation, but often overthink, producing unnecessarily long chains-of-thought that raise cost without improving accuracy. Prior reinforcement learning approaches typically rely on a single outcome reward with trajectory-level length penalties, which cannot distinguish essential from redundant reasoning steps and therefore yield blunt compression. Although recent work incorporates step-level signals, such as offline pruning, supervised data construction, or verifier-based intermediate rewards, reasoning length is rarely treated as an explicit step-level optimization objective during RL. We propose Step-wise Adaptive Penalization (SWAP), a fine-grained framework that allocates length reduction across steps based on intrinsic contribution. We estimate step importance from the model's on-policy log-probability improvement toward the correct answer, then treat excess length as a penalty mass redistributed to penalize low-importance steps more heavily while preserving high-importance reasoning. We optimize with a unified outcome-process advantage within group-relative policy optimization. Extensive experiments demonstrate that SWAP reduces reasoning length by 64.3% on average while improving accuracy by 5.7% relative to the base model.

cs.CL↗

arXiv2025

Wizard of Shopping: Target-Oriented E-commerce Dialogue Generation with Decision Tree Branching

The goal of conversational product search (CPS) is to develop an intelligent, chat-based shopping assistant that can directly interact with customers to understand shopping intents, ask clarification questions, and find relevant products. However, training such assistants is hindered mainly due to the lack of reliable and large-scale datasets. Prior human-annotated CPS datasets are extremely small in size and lack integration with real-world product search systems. We propose a novel approach, TRACER, which leverages large language models (LLMs) to generate realistic and natural conversations for different shopping domains. TRACER's novelty lies in grounding the generation to dialogue plans, which are product search trajectories predicted from a decision tree model, that guarantees relevant product discovery in the shortest number of search conditions. We also release the first target-oriented CPS dataset Wizard of Shopping (WoS), containing highly natural and coherent conversations (3.6k) from three shopping domains. Finally, we demonstrate the quality and effectiveness of WoS via human evaluations and downstream tasks.

cs.CL↗

arXiv2024

Identifying High Consideration E-Commerce Search Queries

In e-commerce, high consideration search missions typically require careful and elaborate decision making, and involve a substantial research investment from customers. We consider the task of identifying High Consideration (HC) queries. Identifying such queries enables e-commerce sites to better serve user needs using targeted experiences such as curated QA widgets that help users reach purchase decisions. We explore the task by proposing an Engagement-based Query Ranking (EQR) approach, focusing on query ranking to indicate potential engagement levels with query-related shopping knowledge content during product search. Unlike previous studies on predicting trends, EQR prioritizes query-level features related to customer behavior, finance, and catalog information rather than popularity signals. We introduce an accurate and scalable method for EQR and present experimental results demonstrating its effectiveness. Offline experiments show strong ranking performance. Human evaluation shows a precision of 96% for HC queries identified by our model. The model was commercially deployed, and shown to outperform human-selected queries in terms of downstream customer impact, as measured through engagement.

cs.IR↗

arXiv2024

Generative Explore-Exploit: Training-free Optimization of Generative Recommender Systems using LLM Optimizers

Recommender systems are widely used to suggest engaging content, and Large Language Models (LLMs) have given rise to generative recommenders. Such systems can directly generate items, including for open-set tasks like question suggestion. While the world knowledge of LLMs enable good recommendations, improving the generated content through user feedback is challenging as continuously fine-tuning LLMs is prohibitively expensive. We present a training-free approach for optimizing generative recommenders by connecting user feedback loops to LLM-based optimizers. We propose a generative explore-exploit method that can not only exploit generated items with known high engagement, but also actively explore and discover hidden population preferences to improve recommendation quality. We evaluate our approach on question generation in two domains (e-commerce and general knowledge), and model user feedback with Click Through Rate (CTR). Experiments show our LLM-based explore-exploit approach can iteratively improve recommendations, and consistently increase CTR. Ablation analysis shows that generative exploration is key to learning user preferences, avoiding the pitfalls of greedy exploit-only approaches. A human evaluation strongly supports our quantitative findings.

cs.CL↗

arXiv2024

Identifying Shopping Intent in Product QA for Proactive Recommendations

Voice assistants have become ubiquitous in smart devices allowing users to instantly access information via voice questions. While extensive research has been conducted in question answering for voice search, little attention has been paid on how to enable proactive recommendations from a voice assistant to its users. This is a highly challenging problem that often leads to user friction, mainly due to recommendations provided to the users at the wrong time. We focus on the domain of e-commerce, namely in identifying Shopping Product Questions (SPQs), where the user asking a product-related question may have an underlying shopping need. Identifying a user's shopping need allows voice assistants to enhance shopping experience by determining when to provide recommendations, such as product or deal recommendations, or proactive shopping actions recommendation. Identifying SPQs is a challenging problem and cannot be done from question text alone, and thus requires to infer latent user behavior patterns inferred from user's past shopping history. We propose features that capture the user's latent shopping behavior from their purchase history, and combine them using a novel Mixture-of-Experts (MoE) model. Our evaluation shows that the proposed approach is able to identify SPQs with a high score of F1=0.91. Furthermore, based on an online evaluation with real voice assistant users, we identify SPQs in real-time and recommend shopping actions to users to add the queried product into their shopping list. We demonstrate that we are able to accurately identify SPQs, as indicated by the significantly higher rate of added products to users' shopping lists when being prompted after SPQs vs random PQs.

cs.CL↗

arXiv2024

Instant Answering in E-Commerce Buyer-Seller Messaging using Message-to-Question Reformulation

E-commerce customers frequently seek detailed product information for purchase decisions, commonly contacting sellers directly with extended queries. This manual response requirement imposes additional costs and disrupts buyer's shopping experience with response time fluctuations ranging from hours to days. We seek to automate buyer inquiries to sellers in a leading e-commerce store using a domain-specific federated Question Answering (QA) system. The main challenge is adapting current QA systems, designed for single questions, to address detailed customer queries. We address this with a low-latency, sequence-to-sequence approach, MESSAGE-TO-QUESTION ( M2Q ). It reformulates buyer messages into succinct questions by identifying and extracting the most salient information from a message. Evaluation against baselines shows that M2Q yields relative increases of 757% in question understanding, and 1,746% in answering rate from the federated QA system. Live deployment shows that automatic answering saves sellers from manually responding to millions of messages per year, and also accelerates customer purchase decisions by eliminating the need for buyers to wait for a reply

cs.CL↗

arXiv2024

Controllable Decontextualization of Yes/No Question and Answers into Factual Statements

Yes/No or polar questions represent one of the main linguistic question categories. They consist of a main interrogative clause, for which the answer is binary (assertion or negation). Polar questions and answers (PQA) represent a valuable knowledge resource present in many community and other curated QA sources, such as forums or e-commerce applications. Using answers to polar questions alone in other contexts is not trivial. Answers are contextualized, and presume that the interrogative question clause and any shared knowledge between the asker and answerer are provided. We address the problem of controllable rewriting of answers to polar questions into decontextualized and succinct factual statements. We propose a Transformer sequence to sequence model that utilizes soft-constraints to ensure controllable rewriting, such that the output statement is semantically equivalent to its PQA input. Evaluation on three separate PQA datasets as measured through automated and human evaluation metrics show that our proposed approach achieves the best performance when compared to existing baselines.

cs.CL↗

arXiv2023

InstructPTS: Instruction-Tuning LLMs for Product Title Summarization

E-commerce product catalogs contain billions of items. Most products have lengthy titles, as sellers pack them with product attributes to improve retrieval, and highlight key product aspects. This results in a gap between such unnatural products titles, and how customers refer to them. It also limits how e-commerce stores can use these seller-provided titles for recommendation, QA, or review summarization. Inspired by recent work on instruction-tuned LLMs, we present InstructPTS, a controllable approach for the task of Product Title Summarization (PTS). Trained using a novel instruction fine-tuning strategy, our approach is able to summarize product titles according to various criteria (e.g. number of words in a summary, inclusion of specific phrases, etc.). Extensive evaluation on a real-world e-commerce catalog shows that compared to simple fine-tuning of LLMs, our proposed approach can generate more accurate product name summaries, with an improvement of over 14 and 8 BLEU and ROUGE points, respectively.

cs.CL↗

arXiv2023

Follow-on Question Suggestion via Voice Hints for Voice Assistants

The adoption of voice assistants like Alexa or Siri has grown rapidly, allowing users to instantly access information via voice search. Query suggestion is a standard feature of screen-based search experiences, allowing users to explore additional topics. However, this is not trivial to implement in voice-based settings. To enable this, we tackle the novel task of suggesting questions with compact and natural voice hints to allow users to ask follow-up questions. We define the task, ground it in syntactic theory and outline linguistic desiderata for spoken hints. We propose baselines and an approach using sequence-to-sequence Transformers to generate spoken hints from a list of questions. Using a new dataset of 6681 input questions and human written hints, we evaluated the models with automatic metrics and human evaluation. Results show that a naive approach of concatenating suggested questions creates poor voice hints. Our approach, which applies a linguistically-motivated pretraining task was strongly preferred by humans for producing the most natural hints.

cs.CL↗

arXiv2023

MultiCoNER v2: a Large Multilingual dataset for Fine-grained and Noisy Named Entity Recognition

We present MULTICONER V2, a dataset for fine-grained Named Entity Recognition covering 33 entity classes across 12 languages, in both monolingual and multilingual settings. This dataset aims to tackle the following practical challenges in NER: (i) effective handling of fine-grained classes that include complex entities like movie titles, and (ii) performance degradation due to noise generated from typing mistakes or OCR errors. The dataset is compiled from open resources like Wikipedia and Wikidata, and is publicly available. Evaluation based on the XLM-RoBERTa baseline highlights the unique challenges posed by MULTICONER V2: (i) the fine-grained taxonomy is challenging, where the scores are low with macro-F1=0.63 (across all languages), and (ii) the corruption strategy significantly impairs performance, with entity corruption resulting in 9% lower performance relative to non-entity corruptions across all languages. This highlights the greater impact of entity noise in contrast to context noise.

cs.CL↗

arXiv2023

Generate-then-Retrieve: Intent-Aware FAQ Retrieval in Product Search

Customers interacting with product search engines are increasingly formulating information-seeking queries. Frequently Asked Question (FAQ) retrieval aims to retrieve common question-answer pairs for a user query with question intent. Integrating FAQ retrieval in product search can not only empower users to make more informed purchase decisions, but also enhance user retention through efficient post-purchase support. Determining when an FAQ entry can satisfy a user's information need within product search, without disrupting their shopping experience, represents an important challenge. We propose an intent-aware FAQ retrieval system consisting of (1) an intent classifier that predicts when a user's information need can be answered by an FAQ; (2) a reformulation model that rewrites a query into a natural question. Offline evaluation demonstrates that our approach improves Hit@1 by 13% on retrieving ground-truth FAQs, while reducing latency by 95% compared to baseline systems. These improvements are further validated by real user feedback, where 71% of displayed FAQs on top of product search results received explicit positive user feedback. Overall, our findings show promising directions for integrating FAQ retrieval into product search at scale.

cs.CL↗

arXiv2023

Answering Unanswered Questions through Semantic Reformulations in Spoken QA

Spoken Question Answering (QA) is a key feature of voice assistants, usually backed by multiple QA systems. Users ask questions via spontaneous speech which can contain disfluencies, errors, and informal syntax or phrasing. This is a major challenge in QA, causing unanswered questions or irrelevant answers, and leading to bad user experiences. We analyze failed QA requests to identify core challenges: lexical gaps, proposition types, complex syntactic structure, and high specificity. We propose a Semantic Question Reformulation (SURF) model offering three linguistically-grounded operations (repair, syntactic reshaping, generalization) to rewrite questions to facilitate answering. Offline evaluation on 1M unanswered questions from a leading voice assistant shows that SURF significantly improves answer rates: up to 24% of previously unanswered questions obtain relevant answers (75%). Live deployment shows positive impact for millions of customers with unanswered questions; explicit relevance feedback shows high user satisfaction.

cs.CL↗

arXiv2023

SemEval-2023 Task 2: Fine-grained Multilingual Named Entity Recognition (MultiCoNER 2)

We present the findings of SemEval-2023 Task 2 on Fine-grained Multilingual Named Entity Recognition (MultiCoNER 2). Divided into 13 tracks, the task focused on methods to identify complex fine-grained named entities (like WRITTENWORK, VEHICLE, MUSICALGRP) across 12 languages, in both monolingual and multilingual scenarios, as well as noisy settings. The task used the MultiCoNER V2 dataset, composed of 2.2 million instances in Bangla, Chinese, English, Farsi, French, German, Hindi, Italian., Portuguese, Spanish, Swedish, and Ukrainian. MultiCoNER 2 was one of the most popular tasks of SemEval-2023. It attracted 842 submissions from 47 teams, and 34 teams submitted system papers. Results showed that complex entity types such as media titles and product names were the most challenging. Methods fusing external knowledge into transformer models achieved the best performance, and the largest gains were on the Creative Work and Group classes, which are still challenging even with external knowledge. Some fine-grained classes proved to be more challenging than others, such as SCIENTIST, ARTWORK, and PRIVATECORP. We also observed that noisy data has a significant impact on model performance, with an average drop of 10% on the noisy subset. The task highlights the need for future research on improving NER robustness on noisy data containing complex entities.

cs.CL↗