The Multilingual Curse at the Retrieval Layer: Evidence from Amharic

Multilingual retrieval increasingly underpins cross-lingual question answering and retrieval-augmented generation. Strong zero-shot scores on multilingual benchmarks are often taken as evidence that current encoders transfer reliably across many languages. We argue that this assumption breaks down for underrepresented, morphologically rich languages, and use Amharic as a diagnostic case. Under a shared passage retrieval protocol covering dense, late-interaction, learned sparse, and cross-encoder paradigms, we compare zero-shot multilingual retrievers, Amharic-fine-tuned multilingual retrievers, and monolingual Amharic retrievers. The strongest zero-shot multilingual retriever underperforms the strongest monolingual Amharic first-stage retriever by 23% relative MRR@10. Fine-tuning two recent multilingual embedding models on the same Amharic supervision yields 32-60% relative MRR@10 gains over zero-shot, but the best Amharic-fine-tuned multilingual model remains below the strongest monolingual Amharic retriever. These findings indicate that zero-shot multilingual retrieval is not a sufficient proxy for equitable information access in the LLM era: for underrepresented languages, retrieval must be evaluated and adapted in-language rather than inferred from aggregate multilingual benchmarks. To foster future research, we publicly release the dataset, codebase, and trained models at https://github.com/rasyosef/amharic-neural-ir.

The Attention Market: Interpreting Online Fair Re-ranking as Manifold Optimization under Walrasian Equilibrium

Fair re-ranking aims to promote long-tail items and enhance diversity within groups in information retrieval. While previous research on online fairness-aware re-ranking has shown promising outcomes, our comprehensive evaluation of online fair re-ranking methods over 20 settings reveals significant performance disparities among existing methods. To uncover the root causes of these inconsistencies, we reformulate fair re-ranking within an attentional market framework governed by a Walrasian Equilibrium, where the fairness is treated as a taxation cost. This market-based formulation is then coupled with manifold optimization, demonstrating that seeking this equilibrium is equivalent to performing gradient descent on a specific ranking manifold constructed by the market. Different re-ranking settings induce distinct manifold geometries, and these intrinsic geometric differences dictate the gradient landscapes and optimization trajectories. We propose ManifoldRank, an efficient online fair re-ranking algorithm. ManifoldRank adjusts gradients to align with the ranking manifold, considering various contextual settings. On the supply side, it incorporates a gradient adjustment based on different fairness requirements, accounting for associated costs. On the demand side, it empirically predicts an additional gradient adjustment term derived from the ranking scores. By integrating these two gradient adjustments, ManifoldRank effectively balances fairness and accuracy. Experimental results across multiple datasets confirm ManifoldRank's effectiveness.

A Parametric Memory Head for Continual Generative Retrieval

Generative information retrieval (GenIR) consolidates retrieval into a single neural model that decodes document identifiers (docids) directly from queries. While this model-as-index paradigm offers architectural simplicity, it is poorly suited to dynamic document collections. Unlike modular systems, where indexes are easily updated, GenIR's knowledge is parametrically encoded in its weights; consequently, standard adaptation methods such as full and parameter-efficient fine-tuning can induce catastrophic forgetting. We show that sequential adaptation improves retrieval on newly added documents but substantially degrades performance on earlier slices, exposing a pronounced stability-plasticity trade-off. To address this, we propose post-adaptation memory tuning (PAMT), a memory-only stabilization stage that augments an adapted model with a modular parametric memory head (PMH). PAMT freezes the backbone and attaches a product-key memory with fixed addressing. During prefix-trie constrained decoding, decoder hidden states sparsely query PMH to produce residual corrections in hidden space; these corrections are mapped to score adjustments via the frozen output embedding matrix, computed only over trie-valid tokens. This guides docid generation while keeping routing and backbone parameters fixed. To limit cross-slice interference, PAMT updates only a fixed budget of memory values selected using decoding-time access statistics, prioritizing entries frequently activated by the current slice and rarely used in prior sessions. Experiments on MS MARCO and Natural Questions under sequential, disjoint corpus increments show that PAMT substantially improves retention on earlier slices with minimal impact on retrieval performance for newly added documents, while modifying only a sparse subset of memory values per session.

Lost in Decoding? Reproducing and Stress-Testing the Look-Ahead Prior in Generative Retrieval

Generative retrieval (GR) ranks documents by autoregressively generating document identifiers. Because many GR methods rely on trie-constrained beam search, they are vulnerable to early pruning of relevant prefixes under finite-beam decoding. Planning Ahead in Generative Retrieval (PAG) mitigates this failure mode by using simultaneous decoding to compute a document-level look-ahead prior that guides subsequent sequential decoding. We reproduce PAG at inference time and stress-test its decoding behavior. Using the authors' released checkpoint and identifier/trie artifacts under the reported decoding setup, we reproduce the main effectiveness results on MS MARCO Dev and TREC-DL 2019/2020, and corroborate the reported beam-size-latency trade-off in our hardware setting. Beyond reproduction, we introduce plan drift diagnostics that quantify how intent-preserving query variations alter the planner's top-n candidate set and highest-weight planner tokens, and how these changes affect guided decoding. We find that PAG's planning signal is brittle under lexical surface-form variation: intent-preserving typos can trigger plan collapse, where the planned candidate pool shifts enough that the look-ahead bonus provides little useful guidance, effectively reverting decoding toward weaker unguided search. We further evaluate fixed-index cross-lingual robustness using non-English mMARCO queries against an English index, and assess query-side mitigation strategies that require no re-indexing; query translation provides the strongest recovery in our setting. Overall, our results confirm PAG's reported effectiveness and the benefit of planning-guided decoding under the released inference setup, while showing that these gains depend on the stability of the planning signal under realistic query variation and query-document mismatch.

Following the Eye-Tracking Evidence: Established Web-Search Assumptions Fail in Carousel Interfaces

Carousel interfaces have been the de-facto standard for streaming media services for over a decade. Yet, there has been very little research into user behavior with such interfaces, which thus remains poorly understood. Due to this lack of empirical research, previous work has assumed that behaviors established in single-list web-search interfaces, such as the F-pattern and the examination hypothesis, also apply to carousel interfaces, for instance when designing click models or evaluation metrics. We analyze a recently-released interaction and examination dataset resulting from an eye-tracking study performed on carousel interfaces to verify whether these assumptions actually hold. We find that (i)~the F-pattern holds only for vertical examination and not for horizontal swiping; additionally, we discover that, when conditioned on a click, user examination follows an L-pattern unique to carousel interfaces; (ii)~click-through-rates conditioned on examination indicate that the well-known examination hypothesis does not hold in carousel interfaces; and (iii)~contrary to the assumptions of previous work, users generally ignore carousel headings and focus directly on the content items. Our findings show that many user behavior assumptions, especially concerning examination patterns, do not transfer from web search interfaces to carousel recommendation settings. Our work shows that the field lacks a reliable foundation on which to build models of user behavior with these interfaces. Consequently, a re-evaluation of existing metrics and click models for carousel interfaces may be warranted.

AdversarialCoT: Single-Document Retrieval Poisoning for LLM Reasoning

Retrieval-augmented generation (RAG) enhances large language model (LLM) reasoning by retrieving external documents, but also opens up new attack surfaces. We study knowledge-base poisoning attacks in RAG, where an attacker injects malicious content into the retrieval corpus, which is then naturally surfaced by the retriever and consumed by the LLM during reasoning. Unlike prior work that floods the corpus with poisoned documents, we propose AdversarialCoT, a query-specific attack that poisons only a single document in the corpus. AdversarialCoT first extracts the target LLM's reasoning framework to guide the construction of an initial adversarial chain-of-thought (CoT). The adversarial document is iteratively refined through interactions with the LLM, progressively exposing and exploiting critical reasoning vulnerabilities. Experiments on benchmark LLMs show that a single adversarial document can significantly degrade reasoning accuracy, revealing subtle yet impactful weaknesses. This study exposes security risks in RAG systems and provides actionable insights for designing more robust LLM reasoning pipelines.

SubSearch: Intermediate Rewards for Unsupervised Guided Reasoning in Complex Retrieval

Large language models (LLMs) are probabilistic in nature and perform more reliably when augmented with external information. As complex queries often require multi-step reasoning over the retrieved information, with no clear or predetermined reasoning path, they remain challenging. Recent approaches train models using reinforcement learning on the model's outcome, showing promise in improving how models handle complex information. We introduce SubSearch, a specialized framework that shifts from outcome-only supervision to intermediate reward signals that incentivize planning high-quality reasoning. Unlike previous work on process reward modeling, which focuses on training a separate reward model with annotated trajectories by either human annotators or large LLM judges, SubSearch directly optimizes the generator using intrinsic process rewards, which we define as internally-derived rewards, eliminating the need for external supervision, and moving towards autonomous information-intensive reasoning. Experiments on seven benchmarks show that rewarding intermediate reasoning steps with intrinsic rewards leads to more robust reasoning traces in both QA and multi-hop QA datasets over using only outcome rewards. SubSearch can help in building reasoning traces that allow agents to better integrate search engines for complex query answering, while offering a data-efficient alternative to supervised process modeling.

Cold-Starts in Generative Recommendation: A Reproducibility Study

Cold-start recommendation remains a central challenge in dynamic, open-world platforms, requiring models to recommend for newly registered users (user cold-start) and to recommend newly introduced items to existing users (item cold-start) under sparse or missing interaction signals. Recent generative recommenders built on pre-trained language models (PLMs) are often expected to mitigate cold-start by using item semantic information (e.g., titles and descriptions) and test-time conditioning on limited user context. However, cold-start is rarely treated as a primary evaluation setting in existing studies, and reported gains are difficult to interpret because key design choices, such as model scale, identifier design, and training strategy, are frequently changed together. In this work, we present a systematic reproducibility study of generative recommendation under a unified suite of cold-start protocols.

Multi-Step Semantic Reasoning in Generative Retrieval

Generative retrieval (GR) models encode a corpus within model parameters and generate relevant document identifiers directly for a given query. While this paradigm shows promise in retrieval tasks, existing GR models struggle with complex queries in numerical contexts, such as those involving semantic reasoning over financial reports, due to limited reasoning capabilities. This limitation leads to suboptimal retrieval accuracy and hinders practical applicability. We propose ReasonGR, a framework designed to enhance multi-step semantic reasoning in numerical contexts within GR. ReasonGR employs a structured prompting strategy combining task-specific instructions with stepwise reasoning guidance to better address complex retrieval queries. Additionally, it integrates a reasoning-focused adaptation module to improve the learning of reasoning-related parameters. Experiments on the FinQA dataset, which contains financial queries over complex documents, demonstrate that ReasonGR improves retrieval accuracy and consistency, indicating its potential for advancing GR models in reasoning-intensive retrieval scenarios.

ERASE -- A Real-World Aligned Benchmark for Unlearning in Recommender Systems

Machine unlearning (MU) enables the removal of selected training data from trained models, to address privacy compliance, security, and liability issues in recommender systems. Existing MU benchmarks poorly reflect real-world recommender settings: they focus primarily on collaborative filtering, assume unrealistically large deletion requests, and overlook practical constraints such as sequential unlearning and efficiency. We present ERASE, a large-scale benchmark for MU in recommender systems designed to align with real-world usage. ERASE spans three core tasks -- collaborative filtering, session-based recommendation, and next-basket recommendation -- and includes unlearning scenarios inspired by real-world applications, such as sequentially removing sensitive interactions or spam. The benchmark covers seven unlearning algorithms, including general-purpose and recommender-specific methods, across nine public datasets and nine state-of-the-art models. We execute ERASE to produce more than 600 GB of reusable artifacts, such as extensive experimental logs and more than a thousand model checkpoints. Crucially, the artifacts that we release enable systematic analysis of where current unlearning methods succeed and where they fall short. ERASE showcases that approximate unlearning can match retraining in some settings, but robustness varies widely across datasets and architectures. Repeated unlearning exposes weaknesses in general-purpose methods, especially for attention-based and recurrent models, while recommender-specific approaches behave more reliably. ERASE provides the empirical foundation to help the community assess, drive, and track progress toward practical MU in recommender systems.