Large language models (LLMs) are increas- ingly deployed as long-horizon conversational agents, motivating growing interest in mem- ory systems. However, existing benchmarks primarily evaluate memory through QA-style probing rather than in-situ conversational usage. We introduce LOCOMO-CONV, a conversa- tional memory benchmark derived from Lo- CoMo with four query styles: dialog, implicit, counterfactual, and composed. Across five rep- resentative memory systems, we evaluate both retrieval recall and end-to-end response qual- ity. Our experiments show that conversational framing exposes substantial retrieval gaps over- looked by QA benchmarks, especially on im- plicit and composed queries, which multi-facet query rewriting narrows for raw-turn mem- ory but not abstractive memory. We further find that strong retrieval does not fully trans- late into response quality, and that implicit queries exhibit silent grounding, where mem- ory improves contextual grounding without ex- plicitly surfacing the gold fact. These results point to reasoning-based memory elaboration as a promising direction, and we release aux- iliary supportive_memory annotations captur- ing conversationally useful context beyond the original gold evidence.
Large language models (LLMs) struggle to classify text into taxonomies with many semantically similar labels, as the distinctions are domain-specific and not captured by pre-training. To handle large label spaces, a common approach retrieves top-$K$ candidate labels by embedding similarity and prompt the LLM to choose among them. However, top-$K$ retrieval reduces the number of candidates but does not help the model tell similar ones apart. When two similar labels both appear as candidates, the model lacks the signal to choose correctly between them. We propose a framework that (1) identifies which label pairs the model struggles to distinguish, (2) expands the candidate set to include confusable labels, and (3) generates targeted rules to differentiate between similar candidates. The framework requires no fine-tuning, and the generated rules transfer to smaller, cheaper models. On three benchmarks (WOS, Flipkart, LEDGAR), our approach improves Macro F1 by up to 10.0pp over retrieval baselines, with smaller models (2B--20B) gaining up to 11.5pp via cross-model transfer.
Customer-service QA in an AI contact center (AICC) runs under deployment constraints that benchmark QA misses: tight voice-hotline latency and a high cost for unsupported or wrong automatic answers. We deploy a system that answers only from a closed set of verified QA units: it returns a retrieved unit verbatim, or routes to clarify, abstain, or handoff. The index is enriched offline by staged linguistic seeding (SLS): a human authors a per-unit world-grounded slot recipe, gpt-4.1-mini renders it into variants, and a light human gate filters them. One methodology is reused across both domains, so inference stays a single retrieval pass with no query-time generation. On held-out query variants from two industrial domains, SLS lifts hybrid R@1 to 0.881/0.930 (+0.27/+0.34), with gains across all five retrievers tested. At the same gpt-4.1-mini generation budget, SLS beats doc2query by +0.20/+0.32, while cross-provenance evaluation provides additional evidence of transfer across generated-query distributions. Verified-unit answering also removes free-form generation's unsupported-content surface (7-13% versus approximately 0%). We report this as an application study, including negative results.
Furkan Yilmaz, Habibe Aleyna Tasdemir, Muhammed Faruk Gozaycs.CL
We previously reported a ModernBERT encoder trained from scratch for Turkish (MoganBERT-TR) and a single-vector embedding model built on top of it (MoganBERT-embed). This work introduces the third model in that lineage: MoganColBERT-TR, a multi-vector retrieval model that, instead of compressing a query or a document into a single vector, represents it at the token level through a 768->128 projection and scores it with MaxSim late interaction. The model is not trained from scratch: the embedding model's encoder is taken as the starting point and adapted to the ColBERT objective with a single-epoch distillation phase. Training data is produced from two sources - title-to-passage pairs carved out of our own pretraining corpus in the character domain and at sentence boundaries, and two Turkish question-based retrieval sets - and is distilled from the soft scores of a cross-encoder teacher (bge-reranker-v2-m3) over one positive and seven mined negatives. We show that in hard negative mining, rank-based skipping alone is insufficient and must be combined with a group mask and a cosine ceiling. Evaluation is carried out with the official pipeline of TurkColBERT, a benchmark built for Turkish late-interaction retrieval (PLAID index, exact MaxSim), on five Turkish BEIR datasets; none of them appears in our training pool, so all five results are clean zero-shot. With 148.9M parameters, MoganColBERT-TR reaches an overall score of 37.36 (35.53 nDCG@100, 31.81 nDCG@10) averaged over the five datasets and finishes second among the five models compared: it outperforms the twice-as-large ColmmBERT-base-TR on four of five datasets and by +3.05 overall, and the benchmark's largest model by +12.30. The gap to the leading model (mLateOn) is concentrated on ArguAna-TR, the dataset with by far the longest queries.
When a speaker refers to a scene that the listener cannot directly see, the listener must decide whether to preserve its current understanding or revise it as new utterances arrive. Many language systems treat local mismatch as a cue for updating: divergence from the current understanding encourages adjustment. Yet conversational understanding may be more conservative, interpreting mismatching evidence relative to prior understanding rather than immediately revising it. We introduce a controlled, data-driven framework for turn-by-turn preserve/revise decisions in dialogue, where competing revision policies are learned under otherwise identical conditions. We compare four theory-driven revision strategies, each reflecting a different assumption about when listeners should preserve or revise. Two findings stand out. First, a mismatch-driven policy that updates solely based on local divergence reacts strongly to mismatch but destabilizes grounding and degrades retrieval. Second, an uncertainty-sensitive policy extends mismatch-based updating with accumulated evidence, preserving coherent understanding while maintaining strong retrieval performance. Surprisingly, coherent understanding emerges from a counterintuitive pattern: local mismatch promotes preservation, whereas accumulated uncertainty promotes revision, suggesting that listeners maintain prior understanding despite local mismatch and revise only when uncertainty sufficiently accumulates. This pattern is consistent with conceptual pact theory.
Aida Usmanova, Zangir Iklassov, Markus Leippold +1cs.AI cs.LG
Automated fact-checking (AFC) systems retrieve evidence and predict claim veracity, yet evaluations omit simple baselines, systems are developed for a single benchmark and cannot be trusted to generalise across domains. No prior work cross-evaluates the full two-stage retrieve-then-verify pipeline across diverse datasets, complementing retrieval-only studies (Thakur et al., 2021) and single-stage benchmarking studies (Calamai et al., 2025). We benchmark nine models, ranging from random and sparse baselines to fine-tuned transformers, zero-shot LLMs, and the two highest-ranked systems from the AVeriTeC 2025 shared task, across four datasets spanning scientific, open-web, and climate domains. Three findings stand out: (1) on ClimateCheck claim-only and fine-tuned models outperform zero-shot LLM and top-performing AVeriTeC 2025 systems, highlighting that noisy evidence can degrade veracity prediction; (2) system rankings are strongly domain- and metric-dependent: the best model on SciFact (macro-F1 0.70) drops to 0.31 on ClimateCheck, while the AVeriTeC 2025 winner and runner-up swap rankings based on evaluation metrics and datasets; (3) replacing retrieved evidence with gold annotations improves veracity accuracy by 14-22 points across models, confirming retrieval remains primary bottleneck. We release code, pre-processed datasets, and all results to support reproducible AFC research.
Furkan Yilmaz, Habibe Aleyna Tasdemir, Muhammed Faruk Gozaycs.CL cs.AI
Turkish encoder models have adopted modern architectures while leaving the pretraining objective fixed at masked language modelling. This paper introduces MoganBert-TR, a 149M-parameter Turkish encoder foundation model trained from scratch on a language-specifically filtered corpus, together with an embedding model derived from it (MoganBert-Embed). MoganBert-TR is trained over 237.3B tokens with a two-stage CLM-to-MLM curriculum: causal language modelling first, masked language modelling for the remainder, with the transition made inside the stable phase of a WSD schedule. In a controlled ablation under an equal step budget, this design outperforms pure MLM by 2.7-3.7x on Turkish MS MARCO retrieval; the measured mechanism is embedding geometry, where a single direction absorbs 28.1% of the variance under pure MLM against 11.9% under the curriculum. Long-context extension and learning-rate decay are then split into two branches after a shared prefix: running the final portion of decay at 1024 context improves the TrGLUE average by 0.49 +/- 0.26 points across five paired seeds (p = 0.013) and beats a model-soup alternative by 0.75 points at ~4.3% additional cost. MoganBert-TR attains 78.41 on TrGLUE, the best among the Turkish ModernBERT models compared, and 77.73 on TabiBench, where it leads two of the eight categories with the largest margin on code retrieval (+3.62 points over TabiBERT). MoganBert-Embed, produced through teacher distillation and multi-signal contrastive fine-tuning, ranks first among student models on the MTEB(Turkish) overall average with 68.30 and reaches 99.5% of its 7.57B-parameter teacher's score with a 51x smaller backbone. The accompanying 50,048-token tokenizer outperforms all compared Turkish tokenizers on compression and fertility across two independent test sets. Weights, tokenizer, embedding model and evaluation code: https://huggingface.co/moganai
We introduce Giga-Embeddings, a family of text embedding models designed to combine strong retrieval quality with efficient serving. Its largest member is a sparse 10B-parameter Mixture-of-Experts encoder with approximately 1.8B active parameters per token. Across English, Russian, multilingual, and code MTEB benchmarks, this model achieves the strongest aggregate performance within the family on all four evaluated suites. In our vLLM benchmark with 1024-token inputs, it processes 114.5k tokens per second, providing 25 percent higher throughput than the dense 3B model and 1.56-2.65x the throughput of the evaluated external systems. The family also includes a dense 3B encoder and a distilled 480M encoder for tighter compute and memory budgets. We train the compact model using a dimension-agnostic objective that aligns teacher and student similarity distributions. The resulting 480M model scores 70.98 on Russian MTEB, surpassing FRIDA while using 42 percent fewer parameters. We release all three model checkpoints.
Fina Polat, Daniel Daza, Pengyu Zhang +2cs.CL cs.AI cs.DB
Entity Disambiguation (ED) is a key task for constructing and using knowledge graphs. State-of-the-art neural approaches commonly model ED as a single task, although it consists of two distinct subproblems: retrieving candidate entities and selecting the correct one given context. Dual-encoder models optimize for both within a shared embedding space, forcing representations to balance high-recall retrieval with fine-grained selection, and they require trained retrievers, which are costly to maintain as knowledge graphs change. While recent work has begun to combine retrievers with LLM-based selectors, the interplay between the two stages has not been studied systematically. In this paper, we present a systematic comparison of retrieval strategies for candidate generation under a shared LLM-based selection stage, combining sparse retrieval (BM25), Web KB search, and a state-of-the-art trained dense retriever with several open- and closed-source LLMs. We show that, once selection is delegated to a capable LLM, training the retriever provides only modest additional value: a fully training-free BM25 retriever paired with an LLM selector reaches a new state of the art on the ZELDA benchmark, raising inKB micro-F1 from 82.3 to 86.3 (+4); pairing the same LLM with a trained dense retriever reaches 88.5. Decoupling retrieval from selection also exposes a limitation of current ED systems: when the correct entity is missing from retrieved candidates, they are forced to predict an incorrect entity. In contrast, our framework allows for abstention when retrieval failure is detected. In an evaluation setting that rewards correct abstentions, the training-free BM25 + LLM pipeline reaches 90.7 F1.
Linear attention keeps a state of fixed size. At long context, many stored items share this state, and interference between them degrades retrieval. Gated DeltaNet-2 (GDN-2), like every delta-rule model before it, derives its erase vector from the key of the current token. However, the interference in its reads is measured through the query, and the erase step cannot reach it. We introduce the Query-derived Erase Direction (QED). QED adds a second erase direction derived from the query and orthogonal to the key. In the fast-weight view, a key-directed delta edit cannot change the key-orthogonal part of a read. It uses the editable part to cancel old-state content measured along the query. It also improves retrieval at every length past the training window, and it about doubles the usable context length on S-NIAH-1.
Adnan El Assadi, Niklas Muennighoff, Jinhyuk Leecs.CL
Should you replace your text-embedding pipeline with a large language model? We answer this with a controlled, cost-aware comparison of ten LLMs across six families and 26 embedding models (118M to 14B parameters) on 37 tasks spanning classification, semantic textual similarity (STS), clustering, pair classification, and retrieval. In aggregate the two paradigms are effectively tied: the best LLM (Gemini 3.1 Pro, 77.6) and the best embedding model (77.2) differ by 0.4 points. Their strengths differ by task: LLMs lead on reasoning-heavy retrieval, embedding models lead on classification, and the two match on clustering, STS, and pair classification. Reaching that parity is expensive. An LLM costs up to 1,431x more than an embedding model of comparable quality (USD 154 vs. USD 0.11 per benchmark pass), and the open LLMs tested process tokens 2.5 to 736x more slowly on the same GPU. Reasoning tokens account for 28 to 81% of LLM inference cost; lower reasoning budgets preserve or improve retrieval quality for most models in our ablation. The Pareto frontier contains the leading embedding models and one LLM, Gemini 3.1 Pro. These results support a division of labour: use embedding models for similarity, classification, and clustering, and reserve LLMs for reasoning-intensive retrieval. Our code, datasets, and results are publicly available at https://github.com/embeddings-benchmark/embedders-dilemma.
Traditional Text-to-SQL research and benchmarks assume a known target database, overlooking settings in which a query must be routed within a large, heterogeneous database collection. We therefore study schema linking in a multi-database setting, where the system must first locate the target database and then construct a compact, SQL-relevant schema for generation. We propose MDB-Link, a hierarchical schema-linking framework that retrieves question-relevant columns from a global index, aggregates retrieval evidence to shortlist databases, and uses a budget-aware large language model (LLM) for database reranking, table selection, and column grounding. With Qwen2.5-14B, MDB-Link outperforms LinkAlign on MMQA, Spider2-Snow, and BIRD-dev in database localization and column selection while producing schema subsets close in size to the gold schemas. Exact match improves from 16.88 to 51.41 on MMQA, 2.50 to 9.17 on Spider2-Snow, and 12.52 to 38.01 on BIRD-dev. MDB-Link also runs faster than LinkAlign and AutoLink, demonstrating the effectiveness of hierarchical schema reduction for downstream SQL generation.
Users of modern platforms repeatedly need summaries of recent dialogue, but the window rarely contains enough context to be interpreted on its own. We formalize this setting as streaming dialogue summarization, where a system must summarize a current window using selective memory from an unbounded history under a fixed budget. We show that the central challenge is not how much history is accessed, but whether memory recovers the evidence that the current window presupposes. We construct a benchmark and evaluation protocol that separately assesses whether memory contains gap-resolving evidence and whether the generated summary reflects it. We propose ReMEMBER, a missing-evidence memory framework that conditions retrieval on unresolved window dependencies and refines retrieved chunks into evidence-dense memory under a fixed budget. Experiments on dialogues with histories up to 160K tokens show that ReMEMBER improves memory recall and gap-resolution completeness over memory construction baselines under the same budget.
Personalized generation systems retrieve user history by request--memory relevance and inject it into the model context. Yet relevant history may concern the wrong preference aspect, duplicate public information, or provide insufficient support. We argue that personal memory should be used only when it adds utility beyond a public-only response. We propose TRACE-Memory, a two-stage framework for selective personalization. Stage 1 queries for user-specific information missing from the request and public context, then retrieves a coverage-oriented candidate pool. Stage 2 admits a compact subset of source-traceable evidence units, or the empty set, according to response-level incremental utility. We progressively train the query-generation and evidence-admission policies through structured SFT initialization, reduced-space stage-wise GRPO warm-up, and nested multi-sample Joint GRPO. Across 4,500 Controlled and Natural tasks from Goodreads, Amazon Reviews, and Reddit, TRACE-Memory consistently outperforms random and lexical memory use, improves over semantic retrieval, remains competitive with frontier-LLM memory pipelines as local generator capacity increases, and conditions evidence admission on public-context sufficiency, supporting selective rather than default personalization.
Fifteen years ago, Low proposed that pun translators should stop searching for equivalent words and instead search for new points of contact between sound and meaning. In this paper, we investigate that idea computationally. We model pun translation as a process of discovery, exploration, and selection. A retrieval system searches semantic and phonological neighborhoods for target-language affordances: sound-meaning bridges that may support new wordplay. Multiple language models then explore these opportunities by generating competing translations, while a multi-perspective generate-and-rank architecture selects among them. Beyond system development, our primary contribution is an analysis of how retrieved affordances propagate through the translation process. We find that generators actively exploit retrieved opportunities, evaluators progressively concentrate around stronger sound-meaning bridges, and exact phonological collisions are selected at disproportionately high rates when available. At the same time, many puns still yield no usable affordances, suggesting that retrieval remains the central bottleneck in computational pun translation. The resulting picture is remarkably close to the process envisioned by Low. Successful pun translation emerges not from preserving source-language words, but from discovering new places in the target language where sound and meaning collide.
Long-term personalized dialogue agents must track user preferences as their personas evolve. Existing memory systems organize past events well, but store personas as flat profiles detached from the events that justify them. This loose coupling leads to the memory-persona validity gap and the persona-aware retrieval gap. We propose PGMem, a heterogeneous persona-memory graph that connects event and persona nodes through typed provenance and evidence edges, keeping each persona signal traceable to the events that support or revise it. At retrieval time, PGMem expands from query-relevant seeds and ranks signals by evidential validity. Across three benchmarks with small language model backbones, PGMem consistently outperforms summary-based, persona-aware, graph-structured, and agentic memory baselines, and improves performance as the context grows. The source code of PGMem is available at https://github.com/wonjunchoi23/pgmem/
On-device in-context learning (ICL) relies on pre-inference retrieval to select demonstrations for useful context before downstream model inference. This retrieval must exploit task-specific information while operating over local memories under limited computation, memory, and data-exposure budgets. We propose Conditional Retrieval Alignment (CoRA), a gradient-free framework that converts a frozen encoder into a task-conditioned retriever using paired candidate inputs and outputs. CoRA selects complementary encoder layers, constructs an output-derived conditioning space from candidate memory, and aligns candidate input representations to this space through closed-form ridge regression. Low-rank factorization then produces a compact retrieval basis where candidate outputs are used only during offline index construction, whereas query-time retrieval requires only the query input and precomputed index. We show that CoRA's rank-constrained basis is the optimal low-rank compression of the output-conditioned fitted representation, and derive an exact two-pass streaming construction that avoids materializing the full fitted matrix. We further extend the framework to multimodal exemplar retrieval by incorporating visual representations into the conditioning and retrieval spaces. Experiments across ten textual datasets and four multimodal benchmarks with Llama-3.2-1B, MobileLLM-Pro, OpenFlamingo-3B, and Qwen3.5-2B, as well as end-to-end Raspberry Pi~5 deployment demonstrate that CoRA supports effective task-conditioned retrieval without retriever fine-tuning, backpropagation, or target-model calls.
Language and embedding models used in RAG systems are conventionally assumed to require large-scale pretraining and explicit grounding supervision. We present B1ade, an efficient RAG architecture comprising two purpose-built components: a compact embedding model and a purpose-built SLM. B1ade-embed, a 335M parameter retrieval model constructed via parameter-free fusion of five pretrained encoders achieves top MTEB scores among sub-500M models with zero additional training, and B1ade-1B, an SLM trained on low-cost GPUs using Group Relative Policy Optimization (GRPO) on 723M tokens (2.2M examples) of curated context-question pairs with rewards that optimize only answer similarity. Our central finding is emergent attribution: despite receiving no explicit supervision for source citation, B1ade-1B cites retrieved passages in 42.4% of responses, exceeding the attribution rate of its training distribution by 5.5 percentage points. This demonstrates that grounding behavior can emerge as an accuracy-maximizing strategy under RL training, without explicit reward engineering. On standard QA benchmarks, B1ade-1B achieves 81.82% on PopQA, 65.8% on PubMedQA, and 51.09% on FEVER. In end-to-end RAG evaluation, B1ade-1B achieves an average score of 0.654 across correctness, completeness, coherence, and faithfulness, a 10.8% improvement over the SFT, while closing the gap with models 1.5x its size. These results show that strategic model composition and reward design suffice for resource-efficient RAG, without large-scale pretraining.
Pruning a long context means committing to the blocks a model will keep, and the usual selector is distilled from a dense teacher's attention. That assumes attention shows which context the answer depends on. We test the assumption on retrieval tasks where the evidence is known exactly, by masking context and measuring whether the answer changes. Attention and causal dependence disagree. Teachers attend to outdated facts that the answer does not depend on, and they attend differently across training runs that use the same evidence. Selectors trained on that attention copy both failures. On a multi-hop retrieval task, a selector distilled from attention routes at 36% to 98% depending on the training run. The same selector trained on causal evidence sets reaches 99% or better on every run. Dense accuracy does not tell the teachers apart. Masking the frozen teacher recovers the causal sets of these tasks without annotations. Frozen pretrained models show the same conflict, and selectors supervised with known evidence labels beat attention-based eviction through 32B when context must be pruned before the question arrives.
Low-rank adaptation introduces a static learned update applied identically to every input. The update provides task-level adaptation but does not explicitly represent token-level or instance-level state variation. A family of adapters is proposed that introduces selective state-space recurrence at two complementary granularities. At the token level, \textbf{MaLoRA} (Mamba-modulated low-rank adaptation) makes the adapter's scaling factor a dynamic input-dependent function with recurrent state across tokens, in contrast to the stateless modulators of prior work. At the context level, \textbf{MaRA} (Mamba Retrieval Adapter) tracks cross-segment state and selects the segments most relevant to the query, before the modulated language model generates its answer. Across three frozen backbones (Qwen-2.5-7B, Llama-3.1-8B, Gemma-2-9B) and two reasoning benchmarks (MuSiQue, 2WikiMultihopQA), the family improves reasoning accuracy on every cell of the $3{\times}2$ grid, by $+6.8$ F1 ($+10.5\%$ relative) on average and up to $+9.3$ F1 ($+18.2\%$ relative) on the hardest cell over the LoRA baseline, and the token-level gains carry to RULER QA-2 under length stress.
In multi-hop RAG evaluation, a top-k answer score can hide two different failures: the retrieval window may drop part of the support chain, or it may contain support in a form the adapted reader does not use well. We call this reader-facing form of retrieved evidence an evidence interface. Using three support-annotated multi-hop QA benchmarks, we compare matched adapted readers trained with raw context, retrieval windows, and gold-support diagnostic renderings. These comparisons distinguish support-availability failures from remaining reader-interface effects. Top-k windows become interpretable only after checking whether the complete annotated support chain survives: when it does, short ranked windows can match or improve over raw context; when it does not, missing support explains much of the loss. Gold support-first improves matched readers; on 2Wiki and MuSiQue, a support-supervised ranker raises coverage and recovers raw-context quality at lower prompt cost, while retaining gold headroom. Support-removal checks further show that the gains rely on exposed evidence, not only answer priors. On support-annotated evaluations, top-k answer scores should therefore be reported together with complete-support coverage.
We describe our BioASQ Task 14B 2026 system. The work centers on two design decisions: how aggressively to re-retrieve when first-stage retrieval is weak, and how to combine multiple language-model answers. Retrieval unions two parallel pipelines - a hybrid first stage (dense BGE + BM25 + RRF, reaching R@200 = 99.3% on the BioASQ-13b historical archive) and an agent-driven pipeline that decomposes the question over PubMed, Europe PMC, and iCite - with a BGE cross-encoder quality gate flagging weakly-supported questions for selective re-retrieval. On Task 12B 2024 validation, a cost-pragmatic re-retrieval policy beats a skill-strict baseline significantly on list F1 and list precision, at 12% lower re-retrieval cost. Holding prompt and model fixed across val and test 13B (different question sets), list F1 rises by +0.132 absolute on the BioASQ-released gold-input pool, consistent with substantial retrieval-side headroom. For Phase B answering we decompose multi-model ensemble lift into a selection component bounded by the per-question oracle and a fusion component that aggregators can exceed. The decomposition predicts before any experiment that LLM-as-judge wins on selection-dominated metrics (yes/no, multi-reference ROUGE) but is structurally insufficient on the recall component of fusion-friendly metrics (factoid rank-1, list recall). On Task 13B 2025 our synonym-union resolver wins list recall on every head, while GPT-5.5 solo retains the list-F1 lead because the resolver's wider item set costs precision. On the Task 14B 2026 preliminary leaderboard our team places first on the combined-exact aggregate on three of the eight (phase x batch) leaderboards, wins four individual question-type cells, and takes #1 on Phase B b3 ideal.
Limited memory language models (LMLMs) externalize factual knowledge during pretraining to a knowledge base (KB), rather than memorizing it in their weights. During generation, the model then fetches knowledge from the KB as needed. This recently introduced paradigm provides multiple advantages, including knowledge control capabilities that remain beyond conventional LLMs. We propose continuous-query LMLM (CO-LMLM), where the KB pairs continuous keys with textual knowledge values, a significant departure from prior reliance on relational KB and queries. CO-LMLM generates flexible vector queries at minimal cost, while still integrating human-readable and attributable retrieved knowledge into its generation. We pair this design with an annotation pipeline that tags free-form factual spans in arbitrary text, removing prior work's restriction to Wikipedia. Across pretraining on Wikipedia and FineWeb-Edu and at multiple model scales, CO-LMLM outperforms prior LMLMs and vanilla LLMs in both perplexity and factual precision. At 360M scale, this includes lower perplexity than models pretrained on 40x more data, and SimpleQA-verified performance that is in line with gpt-4o-mini and higher than Claude Sonnet 4.5.
Mohammad Saifullah, Thomas Kornmaier, Taaha Kazi +3cs.AI cs.CL cs.IR
Long-form fiction writers need memory that answers multi-hop questions about evolving story state: who knows a secret and when they learned it, whether an event preceded the narration that revealed it, whether a setup paid off, and how a relationship shifted. General-purpose retrieval and agent-memory systems represent entities and facts but not the narratological structure these questions turn on, so they surface the wrong evidence or none at all. We introduce the Narrative World Model (NWM), a writer-memory system that pairs a narratology-grounded typed temporal-state graph with query-conditioned hybrid retrieval. To measure memory rather than the answerer, we read every system through a single held-constant Opus 4.8 reader over only that system's chapter-safe evidence, on a reproducible public corpus and a validated multi-hop benchmark, and we compare against the strongest existing temporal-knowledge-graph agent-memory framework, Graphiti/Zep (Rasmussen et al., 2025). NWM substantially and significantly outperforms this baseline on multi-hop narratological QA across both corpora, and far exceeds GraphRAG and flat retrieval. The advantage is representational rather than an artifact of extraction: it survives rebuilding the baseline with NWM's own extractor, and traces to its narratology-grounded structure and query-conditioned retrieval, not to graph size or extractor quality.
Language models (LMs) raise an intriguing alternative to vector-based retrieval: conditioning on an in-context corpus and directly generating a relevant answer. However, prior work has largely focused on proprietary systems or the smaller-scale reranking task, leaving corpus-scale in-context retrieval largely unexplored. In this work, we present the first systematic study of in-context retrieval on two scales practical retrievers demand: million-token corpora and length-generalization far beyond training-time sizes. We first introduce BlockSearch, a 0.6B LM retriever whose architectural and training modifications improve over prior LM baselines and length-generalize up to 10 times beyond its training regime. Nevertheless, retrieval still collapses under more extreme extrapolation. We trace this failure to an attention dilution effect: as the corpus grows, irrelevant documents dominate the softmax denominator, reducing the normalized mass on the gold document even when its pre-softmax score stays high. Motivated by this analysis, we introduce length-aware adjustments to the attention softmax and document-level sparse attention. With these modifications, at the million-token scale, our model matches dense retrieval on widely studied benchmarks (e.g, MS MARCO and NQ), while outperforming the concurrent model MSA despite being 7 times smaller. Furthermore, it significantly outperforms dense retrieval on tasks requiring entirely different notions of similarity, such as LIMIT, achieving a 3 times higher score. Together, our results position in-context retrieval a promising alternative to classical retrieval while emphasizing attention control under extreme context growth as a new challenge.
Mechanistic interpretability has produced a rich inventory of component-level analyses that characterise what neural-network components encode and how they interact. Their outputs, however, are not easily reusable: selectivity tables, circuit diagrams, and feature lists remain locked in per-study notebooks - non-composable, not queryable in natural language, and not directly actionable for downstream audit or intervention. We study the representation layer that sits between these analyses and downstream use as a bottleneck that can be evaluated independently, and introduce Manifestation Units, a typed tuple protocol (E, S, R, D, G) extended with attention-head primitives (T) for transformer architectures, organising per-component statistics into structured fields populated automatically and queried through hybrid retrieval. Instantiated across generative vision (beta-VAE), discriminative vision (CNN), and language (GPT-2), the protocol supports two findings: typed structure substantially outperforms unstructured baselines on retrieval, and CNN filters retrieved by the schema satisfy causal sufficiency and necessity criteria under matched-budget controls. The schema absorbs attention-head primitives without modification, set-recovers known IOI circuit members under retrieval-budget-matched controls, and reveals an irreducible two-field core (S+R) with remaining fields either redundant or actively interfering. We present this as schema infrastructure for mechanistic interpretability rather than frontier-scale validation.
Linrui Ma, Chun Hei Lo, Xinyu Wang +12cs.CL cs.AI cs.LG
The quadratic computational cost of traditional attention mechanisms poses a major bottleneck to the scalability and practical deployment of large language models (LLMs), particularly in long-context scenarios. To improve efficiency, existing approaches often enforce rigid structural constraints such as local attention windows. However, these strategies typically lead to substantial performance degradation on tasks requiring precise long-range recall. In this work, we propose MATCH, a scalable and efficient framework that augments sparsified attention mechanisms with dynamically integrated in-context information through an efficient retrieval system. Empirical results show that MATCH significantly improves the performance of sparse-attention models on both synthetic and real-world natural-language tasks. These findings highlight the versatility of MATCH as a general approach for enhancing in-context retrieval capabilities while maintaining the efficiency benefits of sparse attention architectures.
Conversational memory retrieval optimizes relevance, yet a retrieved memory can be relevant and simultaneously outdated: a later turn updates, corrects, or supersedes it. ConvMemory v3 adds a validity context layer that detects and surfaces this update evidence through target-conditioned relation verification, sitting after the v1/v2 retrieval path. The core mechanism is a dual-evidence gate that conditions a relation judgment on the specific target proposition, scoring a (target, source) pair through the product of a MiniLM slot head and a DeBERTa-v3 slot head and gating it by conservative event/operation evidence. On a synthetic multi-hop validity benchmark the gate reaches 90.12% +/- 1.73 accuracy; through a real-data feedback loop that mines failure patterns but trains on synthetic pairs only, the verifier transfers to Memora role binding with zero target-side labels, reaching 98.8% +/- 0.9 group-all-correct. The deployed layer preserves retrieval by default: a context mode attaches structured validity metadata while keeping the candidate set and rank order fixed, and a query-conditioned demote mode is an explicit opt-in for dense current-state workloads, where it raises current-active H@1 from a never-demote baseline of 45.1% to 95.7% +/- 1.2 while protecting non-superseded memories at 99.4% recall. Six machine-verifiable safety contracts pin the layer's behavior. Multi-hop graph propagation is validated as a mechanism; fully automatic construction of strict prerequisite edges is characterized as a boundary, since strict necessity requires counterfactual world knowledge. This report extends ConvMemory v1 (arXiv:2605.28062) and v2 (arXiv:2606.10842).
A personalized language agent must convert a user's interaction history into behavior on each new request at inference time. Two strategies dominate. Retrieval pulls a few of the user's most relevant past items into the prompt, which is accurate but pays a per-query selection and context cost that grows with the history. Distillation instead compresses the history once into a compact natural-language persona, which is bounded, query-independent, and interpretable, but is widely assumed to sacrifice accuracy. Whether, and on which tasks, a distilled persona can match retrieval has not been characterized cleanly. We introduce PersonaLink, a training-free method that distills a user's history into a bounded three-field persona and recursively refines it: each pass self-evaluates the frozen agent on a held-out slice of the user's own labeled history, rewrites the persona from its errors, and keeps the result only when it does not regress on that slice. Because every comparison shares one frozen 7B backbone and differs only in what is placed in context, the design isolates the effect of representation from that of the model. The result is a clear task-type asymmetry. On 200 users of LaMP-2 (15-way news categorization), PersonaLink reaches 0.745-0.755 accuracy, statistically indistinguishable from BM25 retrieval (0.760-0.765).
LLM-based text embedders have substantially improved retrieval and semantic representation quality, but their deployment remains costly: large backbone models slow down embedding inference, while high-dimensional full-precision embeddings impose substantial storage and bandwidth overhead on large-scale indexes. In this paper, we present BITEMBED, an extreme low-bit framework for LLM-based text embedding that jointly targets encoding efficiency and vector storage. BITEMBED converts pretrained LLM backbones into BitNet-style embedding encoders with ternary weights, quantized activations, and lightweight normalization refinement. The converted model is adapted to representation learning through continual contrastive pre-training, followed by supervised contrastive fine-tuning with both similarity-distribution distillation and attention-relation distillation from a full-precision teacher. Beyond quantizing the backbone, BITEMBED further trains output embeddings to support multiple storage precisions meeting different storage needs in various scenarios. Experiments on MMTEB (eng, v2) with Qwen3-0.6B and Gemma3-270M show that BITEMBED is largely comparable to full precision teacher embedders. Moreover, BITEMBED flexibly obtains text embeddings of various precisions, achieving a trade-off between performance and storage cost.