Factual Recall

Large Vision-Language Models (LVLMs) show strong visualization interpretation, yet it is unclear whether their responses reflect genuine reasoning over visual evidence or factual priors learned during training. Current evaluations mix these two sources, obscuring when correct visual interpretation is overridden by memorized facts. We present a framework that isolates visual correctness from factual correctness, revealing validity limitations in existing visualization literacy assessments. Across three experiments with 15 state-of-the-art LVLMs: (1) several models reach human-level performance on standard tests (VLAT), but this may reflect factual recall rather than visual understanding, while randomized-data tests (reVLAT) underestimate literacy when correct visual interpretation is superseded by factual priors. (2) Using our Counterfactual Visualization Literacy Assessment Test (CVLAT) with capability-normalized arbitration metrics, we classify models by the sign of their visual-factual reliance index (VFRI), revealing a visualization-oriented majority and a factual knowledge-oriented minority, though several near-zero cases warrant caution. A human baseline (N=30) on the same counterfactual items confirms that people overwhelmingly follow the chart under conflict, providing a human reference point. (3) Prompt-based intervention can shift prioritization, but its effectiveness is highly model-dependent and direction-asymmetric, and high chart-reading capability does not predict prompt-controllability. Overall, high visualization accuracy is not sufficient evidence of faithful visual reasoning: reliable integration into visual analytics requires evaluating not only visualization literacy but also how models arbitrate between visual evidence and factual priors when the two diverge. Benchmark and code: https://github.com/JaeyoungKim-HCIL/CVLAT

The CoVR-R challenge evaluates composed video retrieval, where a system must retrieve a target video from a large gallery given a reference video and a textual edit instruction. This setting is not a standard video-text retrieval problem: the query is defined by both the visual evidence in the source video and the transformation implied by the edit. A strong embedding model can provide scalable candidate recall, but it may under-express target-side consequences such as state changes, action replacement, object preservation, or temporal consistency. A pairwise multimodal reranker can verify such details more directly, but exhaustive reranking over the full gallery is computationally infeasible. We present $\mathbb{R}^3$, a zero-shot composed video retrieval pipeline built around Reasoning-guided Recalling and Reranking. The core idea is to turn the source-edit query into a reasoning-grounded retrieval program rather than treating the edit text as a short caption. First, the model generates a reasoning trace that describes the expected target video after applying the edit. Then the trace is encoded together with the source video as a reasoning-augmented query, and its retrieval score is fused with the base composed query through an agreement-gated residual rule. At last, a re-ranker verifies the recalled candidates with direct source-candidate comparison. Experiments have demonstrated the effectiveness of our method in addressing this challenge. Codes are available on https://github.com/Lee-zixu/R-3.

Large language models (LLMs) have seen widespread adoption across various domains, yet their reliability is frequently undermined by hallucinations - responses that are plausible-sounding but factually incorrect. In high-stakes domains, these errors can reduce trust and introduce real-world risk. To address this challenge, we present a parameter-efficient approach that uses soft prompts to mitigate hallucinated content and promote responsible abstention in generative question-answering (QA) tasks. Our method, called Responsible Contrastive Soft Prompting (RCSP), uses a composite loss to train soft prompts that balance three goals: suppressing hallucinatory content, encouraging abstention under uncertainty, and preserving or improving factual recall. To achieve these goals, we incorporate contrastive loss, curriculum learning, and KL regularization into our training mechanism. We evaluate our approach on five diverse generative QA datasets using an LLM-as-a-Judge framework. Experimental results on the Gemma 3 (12B) and Llama 3.1 (8B) backbones demonstrate that RCSP effectively balances factual recall with hallucination suppression and abstention, yielding a generally superior F-score over standard reasoning and instruction-based prompting baselines. Notably, these improvements are achieved by training only a fraction of the parameters required by other tuning techniques. Our results demonstrate that soft prompts provide a modular and computationally efficient path toward improving LLM reliability.

Long-horizon search agents accumulate large amounts of retrieved content across many tool calls, making context-budget efficiency increasingly important. A minimal intervention is to mask stale observations from the context as the trajectory progresses, but it remains unclear when this form of context management helps and why. We study observation masking through a systematic sweep over various agent backbones (4B to 284B parameters) and three retrievers on offline and live-web agentic search benchmarks. We find that the accuracy gain from masking follows an asymmetric inverted-U shape when plotted against the model's accuracy without context management: a plateau under weak retrievers, a peak when a strong retriever meets a mid-capacity model, and a sharp collapse when the model is saturated. This pattern reflects the interaction between retriever recall and the model's implicit filtering capacity, rather than either factor in isolation. Mechanistically, masking implements a token-for-turn trade-off: it removes observations the model has largely stopped attending to and pages the agent rarely re-opens. The added turns help when they convert failures into successes, but they fail when masking removes evidence the model would otherwise have used. We therefore reframe context management as a regime-dependent intervention and provide a holistic perspective for analyzing context use in agentic deep search. We release our scaffold and trajectories here (https://github.com/i-DeepSearch/observation-masking) to support future research.

Uncertainty Quantification is a large and growing subfield of large language model behavioral analysis. Primarily to recognize and combat hallucination, the field has largely focused on measuring and improving calibration, the accuracy of uncertainty judgments to task efficacy. In this work, we investigate the relatively underexplored question of how similar large language model uncertainty is to human uncertainty. We investigate the presence and strength of human-similar uncertainty signals, deemed uncertainty alignment, in large language model overt behavior and internal activation patterns. We identify whether the models show evidence of simultaneous alignment and calibration on a variety of datasets covering both multiple choice and open ended factual recall. And we characterize the effect of instruct fine-tuning on each of these facets.

Large Language Models (LLMs) must continuously learn and update knowledge to remain effective in dynamic real-world environments. While Low-Rank Adaptation (LoRA) is widely used for such memory updates, existing studies mainly rely on qualitative downstream evaluations, leaving the quantitative capacity limits and underlying dynamics of exact parametric memory largely unexplored. To bridge this gap, we employ LoRA as a controlled memory capacity probe within the latent space to systematically quantify exact parametric memory. We introduce the Parametric Memory Law, a robust power law linking loss reduction Delta L to effective parameters and sequence length. At the token level, fine-grained analysis reveals a deterministic phase transition, demonstrating that a prediction probability of p > 0.5 constitutes a sufficient condition for verbatim recall under greedy decoding. Driven by these insights, we introduce MemFT, a threshold-guided optimization strategy that dynamically redistributes the training budget toward sub-threshold tokens. Empirical evaluations demonstrate that MemFT can enhance memory fidelity and efficiency. Code will be released at https://github.com/zjunlp/ParametricMemoryLaw.

End-to-end agent-memory benchmarks report a single hit@k per retriever, confounding lexical leakage (uncontrolled query/gold/distractor entity overlap) with tag-mixing (preferences, services, tools averaged together). We propose entity-collision, a system-agnostic protocol that pins the BM25 floor by construction -- every distractor shares the answer's entity tokens -- and stratifies queries by discriminator tag, so any lift over BM25 is attributable to the embedder. Applied to an open-source agent-memory testbed across 5 tags x 3 embedders x 5 collision degrees with paired-bootstrap 95% CIs, the protocol reveals a two-axis pattern: a 256-d hash trigram helps only on closed-vocabulary lexical tags at deep collision; MiniLM-384 dominates both axes; and a 2.7x-parameter BGE-large does not uniformly improve on MiniLM -- it wins on intent-style queries but loses on lexical ones. Encoder capacity alone is not the binding constraint. The synthetic intent-tag null replicates on LongMemEval (n=500) as a single-session-preference recall cliff. Adaptive vector-weight routing on LoCoMo is a measured null: 11.7pp of oracle headroom exists, but no signal we tested recovers it. All 26 result tables and 37 reproduce scripts are version-controlled and verified by a public registry; the protocol is exercised on a deterministically governed memory testbed (event-sourced decision log, DAG-state-machine schema lifecycle) so every reported CI is reproducible byte-for-byte from the ingest stream.

A long-lived LLM agent, such as OpenClaw, earns its value by acting on a user's preferences and constraints across sessions, not just the current request. Yet today's agents keep what a user volunteers but rarely ask for what stays unspoken, leaving a proactivity gap in long-lived LLM agents: an agent cannot act on a preference it never obtained. As users delegate more of their affairs to agents, the impact of this gap grows. We isolate one concrete, controllable slice of this gap as Ask-to-Remember (ATR): the agent decides whether to ask now for a reusable user preference that the current task does not need but a later session with the same user will. ATR is hard even to evaluate: the right question is underdetermined and its payoff deferred to tasks that may never arise. ATRBench, to the best of our knowledge the first ATR benchmark, makes it measurable by fixing each user's preferences as hidden ground truth, so success demands asking, not recall. Across eight frontier LLM agents, defaults fall at least 62 points below an oracle handed the relevant preference, and prompting closes little of it. Diagnostics identify acquisition as the bottleneck. ATRBench surfaces this proactivity gap in current agents and offers a diagnostic testbed for closing it.

In-context learning \ -- performing tasks based on examples given in the prompt \ -- is an important capability that has emerged in large language models and has received significant attention in both theory and practice. Existing theoretical work often focuses on settings where the learning uses information purely from the prompt. However, many practical instances of in-context learning require the model to retrieve factual knowledge stored in the model's parameters, with the context serving to identify which knowledge is relevant. In this work, we study how in-context learning leverages factual knowledge recall. We formalize this behavior by introducing the \emph{in-context factual recall (IC-recall)} task, where a transformer is provided a context of (subject, answer) pairs generated from a hidden relation, along with a query subject, and must both infer this hidden relation and retrieve the corresponding answer. Factual knowledge is modeled by the transformer having access to a simple pre-constructed MLP associative memory storing (subject, relation, answer) triplets. We analyze the supervised fine-tuning dynamics of a one-layer transformer on IC-recall data and prove that the model successfully performs IC-recall by converging to a particular pairwise attention pattern. This fine-tuning stage requires a very small number of samples \ -- only polylogarithmic in the number of stored knowledge triplets. Experiments verify our theoretical predictions and show that the pairwise attention pattern emerges even when the MLP layer is pretrained instead of constructed.

In recent years, several Speech Language Models (SLMs) that represent speech and written text jointly have been presented. The question then emerges about how model-internal mechanisms are similar and different when operating in the two modalities. We focus on how these systems encode, store, and retrieve factual knowledge, which has previously been investigated for text-only models. To investigate mechanisms behind the storage and recall of factual association in SLMs, we leverage Causal Mediation Analysis, a technique previously applied to text-based models. Initial results using SpiritLM, a multimodal model integrating discrete speech tokens reveal discrepancies between text-to-text and speech-to-text results, suggesting that the emergent mechanisms for factual recall are only partially carried over from the text to the speech modality. These results advance our understanding of how internal mechanisms encode factual associations in SLMs while contributing insights for improving speech-enabled AI systems.

Retrieval-Augmented Generation (RAG) has become foundational for grounding large language models in domain-specific corpora, yet conventional vector-based RAG systems are fundamentally limited in their ability to capture the structured, multi-entity relationships that underpin financial market analysis. This paper presents a comprehensive comparative study of a novel two-hop Graph-RAG architecture versus a standard vector-only baseline for cross-entity financial sentiment analysis. Our system constructs a sentiment-weighted knowledge graph of 59 equity entities from 255 news articles covering 10 major technology stocks, then augments dense retrieval with intensity-filtered graph traversal over INFLUENCES edges to surface relational evidence inaccessible to vector search alone. We evaluate both architectures on 100 grounded queries (30 Direct, 70 Relational) using semantic similarity, entity recall, RAGAS metrics, latency benchmarks, and ablation studies. Graph-RAG achieves a statistically significant improvement in entity recall (+6.4%, p < 0.001, Wilcoxon signed-rank) and delivers substantially more relevant answers for complex multi-entity queries (+11.7% Answer Relevancy), with gains concentrating in relational question types (+16.1%). Critically, these improvements come at no measurable cost to answer quality (delta = +0.001 semantic similarity, Cohen's d = 0.078), with a modest 22.6% increase in mean latency offset by an 80% reduction in latency variance. An ablation study on the graph traversal intensity threshold reveals an inverted-U relationship with answer quality, identifying tau = 0.5 as optimal over the production default of tau = 0.7. These findings characterize a precision-for-coverage trade-off inherent to graph-augmented retrieval and provide actionable architectural guidance for practitioners building RAG systems for multi-entity financial analysis.

While scaling laws govern aggregate large language model performance, no scaling law has linked factual recall to both model size and training-data composition. We evaluated 38 models on over 8,900 scholarly references evaluated by an automated reference verification system. Recall quality follows a sigmoid in the log-linear combination of model parameter count and topic representation in training data. These two variables alone explain 60% of the variance across 16 dense models from four families, rising to 74-94% within individual families. The form matches a superposition-inspired account in which recall is gated by a signal-to-noise ratio: signal strength scales with concept frequency and the noise floor with model capacity.

Credit default prediction is a tabular learning problem with severe class imbalance, heterogeneous features, and tight latency budgets. Tabular Foundation Models (TFMs) approach this problem through in-context learning, which makes their predictions sensitive to how the context window is built. We benchmark four classical models and five TFMs on the Home Credit and Lending Club datasets, varying the context-construction strategy (seven options) and the context size (1K to 50K). On both datasets, the choice of context strategy explains more variance in AUC-ROC than the choice of TFM family: balanced and hybrid sampling add 3 to 4 AUC points over uniform sampling, and the gap exceeds the spread between TFMs. With a balanced context of 5K to 10K examples, the strongest TFMs reach the AUC of classical baselines trained on the full data, while also recovering meaningful default-class recall that default-threshold GBDTs do not. We frame this as evidence that context construction, rather than architecture choice, is the primary deployment lever for TFMs in imbalanced credit-risk settings.

Supervised fine-tuning (SFT) is widely used to inject new knowledge into language models, but it often degrades pretrained capabilities such as reasoning and general-domain performance. We argue this forgetting arises because fine-tuning targets from humans or external systems diverge from the model's autoregressive distribution, forcing the optimizer to imitate low-probability token sequences. To address this problem, we propose MixSD, a simple external-teacher-free method for distribution-aligned knowledge injection. Instead of training on fixed targets, MixSD constructs supervision dynamically by mixing tokens from two conditionals of the base model itself: an expert conditional that observes the injected fact in context, and a naive conditional that reflects the model's original prior. The resulting supervision sequences preserve the factual learning signal while remaining substantially closer to the base model's distribution. We evaluate MixSD on two synthetic corpora that we construct to study factual recall and arithmetic function acquisition in a controlled setting, together with established benchmarks for open-domain factual question answering and knowledge editing. Across multiple model scales and settings, MixSD consistently achieves a better memorization-retention trade-off compared to SFT and on-policy self distillation baselines, retaining up to 100% of the base model's held-out capability while maintaining near-perfect training accuracy, whereas standard SFT retains as little as 1%. We further show that MixSD produces substantially lower-NLL supervision targets under the base model and reduces harmful movement along Fisher-sensitive parameter directions. These results suggest that aligning supervision with the model's native generation distribution is a simple and effective principle for knowledge injection that mitigates catastrophic forgetting.

Long-context and memory systems usually treat personalization as a recall problem. In practice, many failures occur later, when a system commits: it turns noisy hints into hard constraints, drops rare witnesses, forgets downstream obligations, or answers despite infeasibility. We introduce Contract-Bounded Evidence Activation (CBEA) with Lexicographic Commitment Validation (LCV). CBEA activates a bounded evidence set using typed coverage, tail witnesses, and consequence debt; LCV validates structured commitments before prose and routes infeasible states to repair, abstention, or recontract. Across 360 fixtures and three generation backends, CBEA+LCV reaches zero failures within validator scope at 0.49-0.60 availability over attempted runs. Raw and long-context baselines with the same LCV gate reach zero only at 0.003-0.092. A shadow oracle diagnostic marks the limit: CBEA+LCV recalls 0.012 of uncompiled visible facts, while raw recalls 0.53. The result is a bounded operating point: explicit commitment control and 74-75% lower median input payload, not universal memory dominance.

Large language model (LLM) based multi-turn dialogue systems often struggle to track dependencies across non-adjacent turns, undermining both consistency and scalability. As conversations lengthen, essential information becomes sparse and is buried in irrelevant context, while processing the entire dialogue history incurs severe efficiency bottlenecks. Existing solutions either rely on high latency external memory or lose fine-grained details through iterative summarization. In this paper, we propose Self-Recall Thinking (SRT), a framework designed to address long-range contextual dependency and sparse informative signals in multi-turn dialogue. SRT identifies helpful historical turns and uses them to generate contextually appropriate responses, enabling the model to selectively recall and reason over context during inference. This process yields an endogenous reasoning process that integrates interpretable recall steps without external modules. SRT incorporates: (1) Dependency Construction: Generating and converting it into self-recall chains; (2)Capability Initialization: Training to enable reasoning chains with recall tokens capability; (3)Reasoning Improvement: Refining accuracy via verifiable rewards to optimize recall and reasoning for correct answers. Experiments on multiple datasets demonstrate that SRT improves F1 score by 4.7% and reduces end-to-end latency by 14.7% over prior methods, achieving a balance between reasoning latency and accuracy, and outperforming state-of-the-art baselines.

Generative retrieval offers a promising alternative by unifying the fragmented multi-stage retrieval process into a single end-to-end model. However, its practical adoption in industrial e-commerce search remains challenging, given the massive and dynamic product catalogs, strict latency requirements, and the need to align retrieval with downstream ranking goals. In this work, we propose a retrieval framework tailored for real-world recall scenarios, positioning generative retrieval as a recall-stage supplement rather than an end-to-end replacement. Our method, CQ-SID (Category-and-Query constrained Semantic ID), employs category-aware and query-item contrastive learning along with Residual Quantized VAEs to encode items into hierarchical semantic cluster identifiers, significantly reducing beam search complexity. Additionally, we develop EG-GRPO (Expert-Guided Group Relative Policy Optimization), a reinforcement learning approach that aligns generative recall with downstream ranking under sparse rewards by injecting ground-truth samples to stabilize training. Offline experiments on TmallAPP search logs show that CQ-SID achieves up to 26.76% and 11.11% relative gains in semantic and personalized click hitrate over RQ-VAE baselines, while halving beam search size. EG-GRPO further improves multi-objective performance. Online A/B tests confirm gains in GMV (+1.15%) and UCTCVR (+0.40%). The generative recall channel now contributes substantially in production, accounting for over 50.25% of exposures, 58.96% of clicks, and 72.63% of purchases, demonstrating a viable path for deploying generative retrieval in real-world e-commerce systems.

Long-term memory is crucial for agents in specialized web environments, where success depends on recalling interface affordances, state dynamics, workflows, and recurring failure modes. However, existing memory benchmarks for agents mostly focus on user histories, short traces, or downstream task success, leaving open how to directly evaluate whether memory systems effectively internalize environment-specific experience. To address this gap, we introduce LongMemEval-V2 (LME-V2), a benchmark for evaluating whether memory systems can help agents acquire the experience needed to become knowledgeable colleagues in customized environments. LME-V2 contains 451 manually curated questions covering five core memory abilities for web agents: static state recall, dynamic state tracking, workflow knowledge, environment gotchas, and premise awareness. Questions are paired with history trajectories containing up to 500 trajectories and 115M tokens. We use a context gathering formulation: memory systems consume history trajectories and return compact evidence for downstream question answering. We propose a suite of two memory methods: AgentRunbook-R, an efficient RAG-based memory with knowledge pools for raw state observations, events, and strategy notes, and AgentRunbook-C, which stores trajectories as files and invokes a coding agent to gather evidence in an augmented sandbox. Experiments show that AgentRunbook-C achieves the best performance with 72.5% average accuracy, outperforming the strongest RAG baseline (48.5%) and the off-the-shelf coding agent baseline (69.3%). Despite the strong performance gains, coding agent based methods have high latency costs. While AgentRunbook-C advances the accuracy-latency Pareto frontier, substantial room for improvement remains. Together, these results establish LME-V2 as a challenging testbed for developing long-term memory systems for environment experience.

How do transformer language models memorize factual associations? A common view casts internal weight matrices as associative memories over pairs of embeddings, requiring parameter counts that scale linearly with the number of facts. We develop a theoretical and empirical account of an alternative, \emph{geometric} form of memorization in which learned embeddings encode relational structure directly, and the MLP plays a qualitatively different role. In a controlled setting where a single-layer transformer must memorize random bijections from subjects to a shared attribute set, we prove that a logarithmic embedding dimension suffices: subject embeddings encode \emph{linear superpositions} of their associated attribute vectors, and a small MLP acts as a relation-conditioned selector that extracts the relevant attribute via ReLU gating, and not as an associative key-value mapping. We extend these results to the multi-hop setting -- chains of relational queries such as ``Who is the mother of the wife of $x$?'' -- providing constructions with and without chain-of-thought that exhibit a provable capacity-depth tradeoff, complemented by a matching information-theoretic lower bound. Empirically, gradient descent discovers solutions with precisely the predicted structure. Once trained, the MLP transfers zero-shot to entirely new bijections when subject embeddings are appropriately re-initialized, revealing that it has learned a generic selection mechanism rather than memorized any particular set of facts.

Large language models demonstrate remarkable ability in factual recall, yet the fundamental limits of storing and retrieving input--output associations with neural networks remain unclear. We study these limits in a minimal setting: a linear associative memory that maps $p$ input embeddings in $\mathbb{R}^d$ to their corresponding~$d$-dimensional targets via a single layer, requiring each mapped input to be well separated from all other targets. Unlike in supervised classification, this strict separation induces~$p$ constraints per association and produces strong correlations between constraints that make a direct characterisation of the storage capacity difficult. Here, we provide a precise characterisation of this capacity in the following way. We first introduce a decoupled model in which each input has its own independent set of competing outputs, and provide numerical and analytical evidence that this decoupled model is equivalent to the original model in terms of storage capacity, spectra of the learnt weights, and storage mechanism. Using tools from statistical physics, we show that the decoupled model can store up to $p_c \log p_c / d^2 = 1 / 2$ associations, and generalise the computation of $p_c$ to linear two-layer architectures. Our analysis also gives mechanistic insight into how the optimal solution improves over a naïve Hebbian learning rule: rather than boosting input-output alignments with broad fluctuations, the optimal solution raises the correct scores just above the extreme-value threshold set by the competing outputs. These findings give a sharp statistical-physics characterisation of factual storage in linear networks and provide a baseline for understanding the memory capacity of more realistic neural architectures.

Large language models (LLMs) have saturated standard medical benchmarks that test factual recall, yet their ability to perform higher-order reasoning, such as synthesizing evidence from multiple sources, remains critically under-explored. To address this gap, we introduce MedMeta, the first benchmark designed to evaluate an LLM's ability to generate conclusions from medical meta-analyses using only the abstracts of cited studies. MedMeta comprises 81 meta-analyses from PubMed (2018--2025) and evaluates models using two distinct workflows: a Retrieval-Augmented Generation (Golden-RAG) setting with ground-truth abstracts, and a Parametric-only approach relying on internal knowledge. Our evaluation framework is validated by a well-structured analysis showing our LLM-as-a-judge protocol strongly aligns with human expert ratings, as evidenced by high Pearson's r correlation (0.81) and Bland-Altman analysis revealing negligible systematic bias, establishing it as a reliable proxy for scalable evaluation. Our findings underscore the critical importance of information grounding: the Golden-RAG workflow consistently and significantly outperforms the Parametric-only approach across models. In contrast, the benefits of domain-specific fine-tuning are marginal and largely neutralized when external material is provided. Furthermore, stress tests show that all models, regardless of architecture, fail to identify and reject negated evidence, highlighting a critical vulnerability in current RAG systems. Notably, even under ideal RAG conditions, current LLMs achieve only slightly above-average performance (~2.7/5.0). MedMeta provides a challenging new benchmark for evidence synthesis and demonstrates that for clinical applications, developing robust RAG systems is a more promising direction than model specialization alone.

Mechanistic interpretability assumes that circuit analysis becomes harder as models scale. We challenge this assumption by showing that the attention architecture matters more than parameter count. Studying three circuit types across Pythia and Qwen2.5, we find that grouped query attention produces circuits that are far more concentrated and mechanistically stable than standard multi-head attention at comparable scales. The same concentration pattern holds across indirect object identification, induction heads, and factual recall. Within a single architecture family (Qwen2.5), factual recall circuits undergo a discrete phase transition above a critical scale, collapsing to a single bottleneck rather than degrading gradually. These findings suggest that some architectural choices make large models more tractable to study and that interpretability difficulty is not a fixed consequence of model size.

Reinforcement learning (RL) has achieved remarkable success in LLM reasoning, but whether it can also improve direct recall of parametric knowledge remains an open question. We study this question in a controlled zero-shot, one-hop, closed-book QA setting with no chain-of-thought, training only on binary correctness rewards and applying fact-level train-test deduplication to ensure gains reflect improved recall rather than reasoning or memorization. Across three model families and multiple factual QA benchmarks, RL yields ~27% average relative gains, surpassing both training- and inference-time baselines alike. Mechanistically, RL primarily redistributes probability mass over existing knowledge rather than acquiring new facts, moving correct answers from the low-probability tail into reliable greedy generations. Our data-attribution study reveals that the hardest examples are the most informative: those whose answers never appear in 128 pre-RL samples (only ~18% of training data) drive ~83% of the gain, since rare correct rollouts still emerge during training and get reinforced. Together, these findings broaden the role of RL beyond reasoning, repositioning it as a tool for unlocking rather than acquiring latent parametric knowledge.

Extraction at ingestion is the wrong primitive for agent memory: content discarded before the query is known cannot be recovered at retrieval time. We propose True Memory, a six-layer architecture that shifts the center of the system from a storage schema to a multi-stage retrieval pipeline operating over events preserved verbatim. The full system runs as a single SQLite file on commodity CPU with no external database, vector index, graph store, or GPU. On LoCoMo (1,540 questions across 10 multi-session conversations), True Memory Pro reaches 93.0% accuracy (3-run mean) against 61.4% for Mem0, 65.4% for Supermemory, approximately 71% for Zep, and 94.5% for EverMemOS under a matched gpt-4.1-mini answer model. On LongMemEval (500 questions), True Memory Pro reaches 87.8% (3-run mean). On BEAM-1M (700 questions at the 1-million-token scale), True Memory Pro reaches 76.6% (3-run mean), above the prior published result of 73.9% for Hindsight. A 56-configuration ablation shows a 1.3-percentage-point spread within the top-performing configuration family.

Motion forecasting often requires trading interpretability for predictive accuracy. Standard anchor-based architectures rely on opaque latent queries that are highly prone to latent collapse, or naive trajectory sampling that limits multi-modal diversity. We propose an end-to-end differentiable framework that grounds predictions in a comprehensive "motion bank", a structured embedding space of physically realizable trajectories constructed via contrastive learning. Rather than regressing paths from a blank slate, our architecture dynamically retrieves explicit motion priors using a novel Anchor Retrieval Layer. This module adapts orthogonally initialized queries via a Dual-Level Gated Cross-Attention mechanism and executes discrete trajectory selection using a Straight-Through Gumbel-Softmax estimator to preserve continuous gradient flow. The retrieved semantically grounded anchors are then geometrically refined by a DETR-style decoder, optimized jointly with a Winner-Takes-All (WTA) kinematic Gaussian Mixture Model (GMM), a latent diversity penalty, and a soft-min weighted endpoint loss. By strictly conditioning the decoding phase on diverse, interpretable motion primitives, our approach eliminates the "black box" of standard latent queries while achieving competitive multi-modal accuracy on the Argoverse 2 and Waymo Open Motion datasets. Code is available at: https://github.com/abviv/recall2predict

When researchers iteratively refine ideas with large language models, do the models preserve fidelity to the original objective? We introduce DriftBench, a benchmark for evaluating constraint adherence in multi-turn LLM-assisted scientific ideation. Across 2,146 scored benchmark runs spanning seven models from five providers (including two open-weight), four interaction conditions, and 38 research briefs from 24 scientific domains, we find that iterative pressure reliably increases structural complexity and often reduces adherence to original constraints. A restatement probe reveals a dissociation between declarative recall and behavioral adherence, as models accurately restate constraints they simultaneously violate. The knows-but-violates (KBV) rate, measuring constraint non-compliance despite preserved recall, ranges from 8% to 99% across models. Structured checkpointing partially reduces KBV rates but does not close the dissociation, and complexity inflation persists. Human validation against blind raters confirms that the LLM judge under-detects constraint violations, making reported constraint adherence scores conservative. Sensitivity analyses confirm the findings are robust to temperature (0.7 vs.\ 1.0) and pressure type (novelty vs.\ rigor). We release all briefs, prompts, rubrics, transcripts, and scores as an open benchmark.

Persistent AI memory is often reduced to a retrieval problem: store prior interactions as text, embed them, and ask the model to recover relevant context later. This design is useful for thematic recall, but it is mismatched to the kinds of memory that agents need in production: exact facts, current state, updates and deletions, aggregation, relations, negative queries, and explicit unknowns. These operations require memory to behave less like search and more like a system of record. This paper argues that reliable external AI memory must be schema-grounded. Schemas define what must be remembered, what may be ignored, and which values must never be inferred. We present an iterative, schema-aware write path that decomposes memory ingestion into object detection, field detection, and field-value extraction, with validation gates, local retries, and stateful prompt control. The result shifts interpretation from the read path to the write path: reads become constrained queries over verified records rather than repeated inference over retrieved prose. We evaluate this design on structured extraction and end-to-end memory benchmarks. On the extraction benchmark, the judge-in-the-loop configuration reaches 90.42% object-level accuracy and 62.67% output accuracy, above all tested frontier structured-output baselines. On our end-to-end memory benchmark, xmemory reaches 97.10% F1, compared with 80.16%-87.24% across the third-party baselines. On the application-level task, xmemory reaches 95.2% accuracy, outperforming specialised memory systems, code-generated Markdown harnesses, and customer-facing frontier-model application harnesses. The results show that, for memory workloads requiring stable facts and stateful computation, architecture matters more than retrieval scale or model strength alone.

Although precise recall is a core objective in Retrieval-Augmented Generation (RAG), a critical oversight persists in the field: improvements in retrieval performance do not consistently translate to commensurate gains in downstream reasoning. To diagnose this gap, we propose the Recall Conversion Rate (RCR), a novel evaluation metric to quantify the contribution of retrieval to reasoning accuracy. Our quantitative analysis of mainstream RAG methods reveals that as Recall@5 improves, the RCR exhibits a near-linear decay. We identify the neglect of retrieval quality in these methods as the underlying cause. In contrast, approaches that focus solely on quality optimization often suffer from inferior recall performance. Both categories lack a comprehensive understanding of retrieval quality optimization, resulting in a trade-off dilemma. To address these challenges, we propose comprehensive retrieval quality optimization criteria and introduce the NeocorRAG framework. This framework achieves holistic retrieval quality optimization by systematically mining and utilizing Evidence Chains. Specifically, NeocorRAG first employs an innovative activated search algorithm to obtain a refined candidate space. Then it ensures precise evidence chain generation through constrained decoding. Finally, the retrieved set of evidence chains guides the retrieval optimization process. Evaluated on benchmarks including HotpotQA, 2WikiMultiHopQA, MuSiQue, and NQ, NeocorRAG achieves SOTA performance on both 3B and 70B parameter models, while consuming less than 20% of tokens used by comparable methods. This study presents an efficient, training-free paradigm for RAG enhancement that effectively optimizes retrieval quality while maintaining high recall. Our code is released at https://github.com/BUPT-Reasoning-Lab/NeocorRAG.

Achieving realistic human-like conversation for virtual characters requires not only a simple memorization and recall of past events, but also the strategic utilization of memory to meet factual needs and social engagement. Current memory utilization relevant (e.g., memory-augmented generation, long-term dialogue, and etc.) benchmarks overlook this nuance, treating memory primarily as a static repository of facts rather than a dynamic resource to be strategically deployed in dialogues. To address this gap, we design StratMem-Bench, a new benchmark to evaluate strategic memory use in character-centric dialogues. This dataset comprises 657 instances where virtual characters must navigate heterogeneous memory pools containing required, supportive, and irrelevant memories. We also propose a framework with different evaluation metrics including Strict Memory Compliance, Memory Integration Quality, Proactive Enrichment Score and Conditional Irrelevance Rate, to evaluate strategic memory use capabilities of virtual characters. Experiments on StratMem-Bench which leverage the state-of-the-art large language models as virtual characters show that all models perform well at distinguishing between required and irrelevant memories, but struggle once supportive memories are introduced into the decision process.

Closed-source frontier labs do not disclose parameter counts. Storing F facts requires at least F/(bits per parameter) weights, so factual recall lower-bounds parameter count--an intrinsic, serving-independent signal, though (as we show) a coarse one. We introduce Incompressible Knowledge Probes (IKPs), a benchmark of 1,400 factual questions spanning 7 tiers of obscurity, designed to isolate knowledge that cannot be derived by reasoning or compressed by architectural improvements. We score with no hallucination penalty (lambda = 0: IKP accuracy is simply the fraction of probed facts answered correctly), which removes both the penalty hyperparameter and the per-tier flooring choice; a full lambda x flooring ablation shows the calibration is robust across scoring choices while individual estimates are not, motivating the no-penalty default. We calibrate a log-linear mapping from IKP accuracy to parameter count on 93 open-weight models (135M-1,600B) spanning 19 vendors, achieving R^2 = 0.910; leave-one-out cross-validation confirms generalization (median fold error 1.48x, 72% within 2x, 86% within 3x). The instrument is deliberately coarse--its 90% prediction interval spans ~3x in either direction, wider than inference economics--so IKP recovers order-of-magnitude effective capacity and relative rankings, not precise parameter counts. For Mixture-of-Experts models, total parameters predict knowledge (R^2 = 0.67) better than active parameters (R^2 = 0.41). We evaluate 201 models from 27 vendors on a curated probe set (1,311 of 1,400 probes surviving name-collision and label-ambiguity filters) and report effective knowledge capacity for all major proprietary frontier models as prediction bands rather than point estimates; for heavily safety-tuned models these are lower bounds, since refusal policy can suppress tens of percentage points of otherwise-answerable capacity.

As the complexity of System-on-Chip (SoC) designs grows, the shift-left paradigm necessitates the rapid development of high-fidelity reference models (typically written in SystemC) for early architecture exploration and verification. While Large Language Models (LLMs) show promise in code generation, their application to hardware modeling faces unique challenges: (1) Rigid, static workflows fail to adapt to varying design complexity, causing inefficiency; (2) Context window overflow in multi-turn interactions leads to catastrophic forgetting of critical specifications; and (3) the Coupled Validation Failure problem--where generated Testbenches (TBs) incorrectly validate flawed models due to correlated hallucinations--severely undermines reliability. To address these limitations, we introduce RefEvo, a dynamic multi-agent framework designed for agile and reliable reference modeling. RefEvo features three key innovations: (1) A Dynamic Design Planner that autonomously decomposes design specifications and constructs tailored execution workflows based on semantic complexity; (2) A Co-Evolutionary Verification Mechanism, which employs a Dialectical Arbiter to simultaneously rectify the model and verification logic against the specification (Spec) oracle, effectively mitigating false positives; and (3) A Spec Anchoring Strategy for lossless context compression. Evaluated on a diverse benchmark of 20 hardware modules, RefEvo achieves a 95% pass rate, outperforming static baselines by a large margin. Furthermore, our context optimization reduces token consumption by an average of 71.04%, achieving absolute savings of over 70,000 tokens per session for complex designs while maintaining 100% specification recall.

Hypernetwork-based methods such as Doc-to-LoRA internalize a document into an LLM's weights in a single forward pass, but they fail systematically on conflicts: when the document contradicts pretraining knowledge, accuracy collapses to 46.4% on the deepest facts. We show the failure is a magnitude problem rather than a representational one. The hypernetwork already targets the right layers, but its adapter margin is approximately constant across documents while the pretrained margin grows with training frequency, so deep conflicts lose by construction. The account predicts that failure should track prior strength: sorting 194 conflicts by the base model's log-probability on the contradicted fact, baseline accuracy falls from 68% on weak-prior questions to 16% on strong-prior ones, a 52 percentage-point gap. The cure is amplitude. Selective Layer Boosting scales the adapter at its top-norm layers, and Conflict-Aware Internalization triggers boosting only when the base model is confident. Both are training-free; together they raise deep-conflict accuracy from 46.4% to 71.0% on Gemma-2B and from 53.6% to 72.5% on Mistral-7B while preserving novel-knowledge recall, and beat vanilla retrieval-augmented generation on medium conflicts by 18 percentage points despite operating entirely in parameter space. We release KID-Bench, a 489-question benchmark that separates novel recall, cross-knowledge combination, and prior-graded conflicts.

Reasoning in large language models is often discussed as a single capability, but some of its gains may stem from simpler underlying operations. We examine two such primitives, recall and state-tracking, through five controlled task families centered on state-based recall, and compare matched transformer and hybrid architectures with and without reasoning augmentation. Across the suite, reasoning-augmented variants substantially outperform instruction-only variants, often by large margins. This pattern is consistent with the State over Tokens view: externalized reasoning traces help because they carry the intermediate state forward in token space. By contrast, hybrid inductive bias does not yield a uniform advantage in accuracy once reasoning tokens are available. When architectural differences do appear, they follow task structure: the hybrid Think model is more robust on strictly sequential chained updates, whereas the transformer Think model is more robust on flat multi-hop retrieval. We therefore cast the main contribution of this study as a descriptive account of what drives performance on state-based recall tasks: reasoning-token augmentation appears to be the dominant factor, while hybrid advantages are narrower, task-dependent, and potentially more about inference efficiency than overall capability. We also release the codebase and data required to reproduce these results.

We introduce Semantic Recall, a novel metric to assess the quality of approximate nearest neighbor search algorithms by considering only semantically relevant objects that are theoretically retrievable via exact nearest neighbor search. Unlike traditional recall, semantic recall does not penalize algorithms for failing to retrieve objects that are semantically irrelevant to the query, even if those objects are among their nearest neighbors. We demonstrate that semantic recall is particularly useful for assessing retrieval quality on queries that have few relevant results among their nearest neighbors-a scenario we uncover to be common within embedding datasets. Additionally, we introduce Tolerant Recall, a proxy metric that approximates semantic recall when semantically relevant objects cannot be identified. We empirically show that our metrics are more effective indicators of retrieval quality, and that optimizing search algorithms for these metrics can lead to improved cost-quality tradeoffs.

Personalized agents that interact with users over long periods must maintain persistent memory across sessions and update it as circumstances change. However, existing benchmarks predominantly frame long-term memory evaluation as fact retrieval from past conversations, providing limited insight into agents' ability to consolidate memory over time or handle frequent knowledge updates. We introduce Memora, a long-term memory benchmark spanning weeks to months long user conversations. The benchmark evaluates three memory-grounded tasks: remembering, reasoning, and recommending. To ensure data quality, we employ automated memory-grounding checks and human evaluation. We further introduce Forgetting-Aware Memory Accuracy (FAMA), a metric that penalizes reliance on obsolete or invalidated memory when evaluating long-term memory. Evaluations of four LLMs and six memory agents reveal frequent reuse of invalid memories and failures to reconcile evolving memories. Memory agents offer marginal improvements, exposing shortcomings in long-term memory for personalized agents.

We study how large language models recall relational knowledge during text generation, with a focus on identifying latent representations suitable for relation classification via linear probes. Prior work shows how attention heads and MLPs interact to resolve subject, predicate, and object, but it remains unclear which representations support faithful linear relation classification and why some relation types are easier to capture linearly than others. We systematically evaluate different latent representations derived from attention head and MLP contributions, showing that per-head attention contributions to the residual stream are comparatively strong features for linear relation classification. Feature attribution analyses of the trained probes, as well as characteristics of the different relation types, reveal clear correlations between probe accuracy and relation specificity, entity connectedness, and how distributed the signal on which the probe relies is across attention heads. Finally, we show how token-level feature attribution of probe predictions can be used to reveal probe behavior in further detail.

Long-horizon enterprise agents make high-stakes decisions (loan underwriting, claims adjudication, clinical review, prior authorization) under lossy memory, multi-step reasoning, and binding regulatory constraints. Current evaluation reports a single task-success scalar that conflates distinct failure modes and hides whether an agent is aligned with the standards its deployment environment requires. We propose that long-horizon decision behavior decomposes into four orthogonal alignment axes, each independently measurable and failable: factual precision (FRP), reasoning coherence (RCS), compliance reconstruction (CRR), and calibrated abstention (CAR). CRR is a novel regulatory-grounded axis; CAR is a measurement axis separating coverage from accuracy. We exercise the decomposition on a controlled benchmark (LongHorizon-Bench) covering loan qualification and insurance claims adjudication with deterministic ground-truth construction. Running six memory architectures, we find structure aggregate accuracy cannot see: retrieval collapses on factual precision; schema-anchored architectures pay a scaffolding tax; plain summarization under a fact-preservation prompt is a strong baseline on FRP, RCS, EDA, and CRR; and all six architectures commit on every case, exposing a decisional-alignment axis the field has not targeted. The decomposition also surfaced a pre-registered prediction of our own, that summarization would fail factual recall, which the data reversed at large magnitude, an axis-level reversal aggregate accuracy would have hidden. Institutional alignment (regulatory reconstruction) and decisional alignment (calibrated abstention) are under-represented in the alignment literature and become load-bearing once decisions leave the laboratory. The framework transfers to any regulated decisioning domain via two steps: build a fact schema, and calibrate the CRR auditor prompt.

While self-consistency is known to improve performance on symbolic reasoning, its effect on the recall of encyclopedic knowledge is unclear due to a lack of targeted evaluation grounds. To address this, we establish such a knowledge recall split for the popular MMLU benchmark by applying a data-driven heuristic from prior work. We validate this split by showing that the performance patterns on the symbolic reasoning and knowledge recall subsets mirror those of GSM8K and MedMCQA, respectively. Using this solid ground, we find that self-consistency consistently improves performance across both symbolic reasoning and knowledge recall, even though its underlying CoT prompting is primarily effective for symbolic reasoning. As a result, we achieve an 89% accuracy on MMLU, the best performance to date with the use of GPT-4o.

Accurate disease classification from radiology reports is essential for many applications. While supervised fine-tuning (SFT) of lightweight LLMs improves accuracy, it can degrade reasoning. We propose a two-stage approach: SFT on disease labels followed by Group Relative Policy Optimization (GRPO) to refine predictions by optimizing accuracy and format without reasoning supervision. Across three radiologist-annotated datasets, SFT outperformed baselines and GRPO further improved classification and enhanced reasoning recall and comprehensiveness.

Robotic manipulation tasks exhibit varying memory requirements, ranging from Markovian tasks that require no memory to non-Markovian tasks that depend on historical information spanning single or multiple interaction trials. Surprisingly, simply extending observation histories of a visuomotor policy often leads to a significant performance drop due to distribution shift and overfitting. To address these issues, we propose Gated Memory Policy (GMP), a visuomotor policy that learns both when to recall memory and what to recall. To learn when to recall memory, GMP employs a learned memory gate mechanism that selectively activates history context only when necessary, improving robustness and reactivity. To learn what to recall efficiently, GMP introduces a lightweight cross-attention module that constructs effective latent memory representations. To further enhance robustness, GMP injects diffusion noise into historical actions, mitigating sensitivity to noisy or inaccurate histories during both training and inference. On our proposed non-Markovian benchmark MemMimic, GMP achieves a 30.1% average success rate improvement over long-history baselines, while maintaining competitive performance on Markovian tasks in RoboMimic. All code, data and in-the-wild deployment instructions are available on our project website https://gated-memory-policy.github.io/.

Retrieval-augmented generation (RAG) remains unreliable in long-form settings, where retrieved evidence is noisy or contradictory, making it difficult for RAG pipelines to maintain factual consistency. Existing approaches focus on retrieval expansion or verification during generation, leaving conflict resolution entangled with generation. To address this limitation, we propose ArbGraph, a framework for pre-generation evidence arbitration in long-form RAG that explicitly resolves factual conflicts. ArbGraph decomposes retrieved documents into atomic claims and organizes them into a conflict-aware evidence graph with explicit support and contradiction relations. On top of this graph, we introduce an intensity-driven iterative arbitration mechanism that propagates credibility signals through evidence interactions, enabling the system to suppress unreliable and inconsistent claims before final generation. In this way, ArbGraph separates evidence validation from text generation and provides a coherent evidence foundation for downstream long-form generation. We evaluate ArbGraph on two widely used long-form RAG benchmarks, LongFact and RAGChecker, using multiple large language model backbones. Experimental results show that ArbGraph consistently improves factual recall and information density while reducing hallucinations and sensitivity to retrieval noise. Additional analyses show that these gains are evident under conflicting or ambiguous evidence, highlighting the effectiveness of evidence-level conflict resolution for improving the reliability of long-form RAG. The implementation is publicly available at https://github.com/1212Judy/ArbGraph.

Users often omit essential details in their requests to LLM-based agents, resulting in under-specified inputs for tool use. This poses a fundamental challenge for tool-augmented agents, as API execution typically requires complete arguments, highlighting the need for personalized tool calling. To study this problem in a more realistic setup, we present Multi-Session Personalized Tool Calling (MPT), a benchmark comprising 4,695 instances over 459 multi-session interaction histories that cover three challenges: Preference Recall, Induction, and Transfer. We further propose PRefine, a test-time memory method that maintains the user's latent preference as a textual hypothesis revised through a generate-verify-refine loop. Across five LLMs, existing memory systems underperform full-history prompting; PRefine outperforms all baselines and alone surpasses it on Preference Transfer. These results indicate that memory for personalized agents must abstract behavior into preferences, rather than simply archive it.

Looped transformers promise test-time compute scaling by spending more iterations on harder problems, but it remains unclear which architectural choices let them extrapolate to harder problems at test time rather than memorize training-specific solutions. We introduce a fixed-point based framework for analyzing looped architectures along three axes of stability -- reachability, input-dependence, and geometry -- and use it to characterize when fixed-point iteration yields meaningful predictions. Theoretically, we prove that looped networks without recall have countable fixed points and cannot achieve strong input-dependence at any spectral regime, while recall combined with outer normalization reliably produces a regime in which fixed points are simultaneously reachable, locally smooth in the input, and supported by stable backpropagation. Empirically, we train single-layer looped transformers on chess, sudoku, and prefix-sums and find that downstream performance tracks the framework's predictions across tasks and architectural configurations. We additionally introduce internal recall, a novel recall placement variant, and show that it becomes competitive with -- and on sudoku, substantially better than -- standard recall placement once outer normalization is applied.

A key capability of modern neural networks is their capacity to simultaneously learn underlying rules and memorize specific facts or exceptions. Yet, theoretical understanding of this dual capability remains limited. We introduce the Rules-and-Facts (RAF) model, a minimal solvable setting that enables precise characterization of this phenomenon by bridging two classical lines of work in the statistical physics of learning: the teacher-student framework for generalization and Gardner-style capacity analysis for memorization. In the RAF model, a fraction $1 - \varepsilon$ of training labels is generated by a structured teacher rule, while a fraction $\varepsilon$ consists of unstructured facts with random labels. We characterize when the learner can simultaneously recover the underlying rule - allowing generalization to new data - and memorize the unstructured examples. Our results quantify how overparameterization enables the simultaneous realization of these two objectives: sufficient excess capacity supports memorization, while regularization and the choice of kernel or nonlinearity control the allocation of capacity between rule learning and memorization. Experiments on a CIFAR10-based RAF task show that the qualitative trade-off persists for fixed kernels, whereas representations learned by deeper neural networks substantially mitigate it, preserving factual recall while improving generalization. The RAF model provides a theoretical foundation for understanding how modern neural networks can infer structure while storing rare or non-compressible information.

Frontier LLM companies have repeatedly assured courts and regulators that their models do not store copies of training data. They further rely on safety alignment strategies via RLHF, system prompts, and output filters to block verbatim regurgitation of copyrighted works, and have cited the efficacy of these measures in their legal defenses against copyright infringement claims. We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task naturally suited for commercial writing assistants, we cause GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 to reproduce up to 85-90% of held-out copyrighted books, with single verbatim spans exceeding 460 words, using only semantic descriptions as prompts and no actual book text. This extraction generalizes across authors: finetuning exclusively on Haruki Murakami's novels unlocks verbatim recall of copyrighted books from over 30 unrelated authors. The effect is not specific to any training author or corpus: random author pairs and public-domain finetuning data produce comparable extraction, while finetuning on synthetic text yields near-zero extraction, indicating that finetuning on individual authors' works reactivates latent memorization from pretraining. Three models from different providers memorize the same books in the same regions ($r \ge 0.90$), pointing to an industry-wide vulnerability. Our findings offer compelling evidence that model weights store copies of copyrighted works and that the security failures that manifest after finetuning on individual authors' works undermine a key premise of recent fair use rulings, where courts have conditioned favorable outcomes on the adequacy of measures preventing reproduction of protected expression.

Enabling artificial intelligence systems, particularly large language models, to update knowledge and flexibly apply it during reasoning remains a central challenge. Existing knowledge editing approaches emphasize atomic facts, improving factual recall but often failing to integrate updated information into a coherent framework usable across contexts. In this work, we argue that knowledge update is fundamentally a reasoning problem rather than a memorization problem. Consequently, a model should be trained in situations where the new information is instrumental to solving a task, combined with pre-existing knowledge, and exercised through multi-step reasoning. Based on this insight, we propose a training strategy based on three principles. First, new knowledge is introduced as a coherent background story that contextualizes novel facts and explains their relation to existing knowledge. Second, models are trained using self-generated multi-hop questions that require multi-step reasoning involving the new information. Third, training is done using knowledge distillation, forcing a student model to internalize the teacher's reasoning behavior without access to the novel information. Experiments show that models trained with this strategy effectively leverage newly acquired knowledge during reasoning and achieve remarkable performance on challenging questions that require combining multiple new facts.

Language models used in retrieval-augmented settings must arbitrate between parametric knowledge stored in their weights and contextual information in the prompt. This work presents a mechanistic study of that choice by extracting an \emph{arbitration vector} from model activations on a curated dataset designed to disentangle (i) irrelevant contexts that elicit parametric recall and (ii) relevant but false contexts that elicit copying. The vector is computed as the residual-stream centroid difference between these regimes across 27 relations, and is injected as an additive intervention at selected layers and token spans to steer behavior in two directions: Copy$\rightarrow$Recall (suppressing context use) and Recall$\rightarrow$Copy (inducing the model to copy any token from the context). Experiments on three architectures (decoder-only and encoder/decoder) and two open-domain QA benchmarks show consistent behavior shifts under moderate scaling while monitoring accuracy and fluency. Mechanistic analyses of attention routing, MLP contributions, and layer-wise probability trajectories reveal an asymmetry: inducing copying is an easy reactivation'' process that can be triggered at different locations in the input, while restoring recall is a suppression'' process that is more fragile and strongly tied to object-token interventions.

An associative memory (AM) enables cue-response recall, and it has recently been recognized as a key mechanism underlying modern neural architectures such as Transformers. In this work, we introduce the concept of distributed dynamic associative memory (DDAM), which extends classical AM to settings with multiple agents and time-varying data streams. In DDAM, each agent maintains a local AM that must not only store its own associations but also selectively memorize information from other agents based on a specified interest matrix. To address this problem, we propose a novel tree-based distributed online gradient descent algorithm, termed DDAM-TOGD, which enables each agent to update its memory on the fly via inter-agent communication over designated routing trees. We derive rigorous performance guarantees for DDAM-TOGD, proving sublinear static regret in stationary environments and a path-length dependent dynamic regret bound in non-stationary environments. These theoretical results provide insights into how communication delays and network structure impact performance. Building on the regret analysis, we further introduce a combinatorial tree design strategy that optimizes the routing trees to minimize communication delays, thereby improving regret bounds. Numerical experiments demonstrate that the proposed DDAM-TOGD framework achieves superior accuracy and robustness compared to representative online learning baselines such as consensus-based distributed optimization, confirming the benefits of the proposed approach in dynamic, distributed environments.

Reinforcement learning (RL) is often credited with improving language model reasoning at the expense of knowledge. We challenge this narrative by showing that reasoning models consistently outperform their instruction-tuned versions on pure knowledge recall tasks. These gains do not reflect newly acquired information, but rather an improved procedural skill in navigating and searching existing knowledge hierarchies within the model parameters. Structured prompting, which explicitly guides models through hierarchical traversal -- recovers most of the instruct-reasoning gap across five model families. A controlled RL experiment on unseen, non-extractable facts improves recall of held-out frequent but previously inaccessible facts, ruling out simple data exposure. On depth-stratified retrieval tasks, reasoning models exhibit superior traversal as retrieval depth grows. Layerwise activation analysis further shows that while factual representations maintain high cosine similarity between instruct and reasoning models, query representations diverge noticeably, indicating that reasoning primarily reshapes how models traverse knowledge rather than the knowledge representation itself. Finally, we find that distilled models often fail to match reasoning models on knowledge recall because they imitate self-correction without acquiring the exploratory behavior needed for hierarchical navigation. Together, these findings suggest that improving factual recall in LLMs depends not only on expanding what models know but also on teaching them to navigate it -- motivating future post-training methods that optimize traversal.

Large language models (LLMs) have shown strong performance in automated code generation, with few-shot prompting widely used for its simplicity and effectiveness. However, few-shot methods depend on curated or manually crafted reference examples, limiting their applicability in data-free coding scenarios such as real-world data-free coding scenarios and benchmarks without training sets. Existing methods that generate reference examples via recitation or analogy cannot guarantee their authenticity or accuracy. Inspired by human metamemory, we propose a novel metamemory agent to enhance one-time code generation in data-free coding scenarios. The agent guides LLMs to recall relevant prior knowledge, evaluate confidence in recalled information, and selectively exploit reliable content for problem solving. This agent removes the need for external reference examples, improves the authenticity and accuracy of recalled knowledge, and adaptively tailors the recall&evaluation process to each task. Extensive experiments demonstrate that the proposed metamemory agent significantly improves one-time code generation quality across data-free coding scenarios. The AI contribution is the metamemory agent, which makes self-recalled examples reliable through confidence evaluation and selection; the engineering application is data-free automated code generation, validated on eight public benchmarks.

Modern LLM-based agents and chat assistants rely on long-term memory frameworks to store reusable knowledge, recall user preferences, and augment reasoning. As researchers create more complex memory architectures, it becomes increasingly difficult to analyze their capabilities and guide future memory designs. Most long-term memory benchmarks focus on simple fact retention, multi-hop recall, and time-based changes. While undoubtedly important, these capabilities can often be achieved with simple retrieval-augmented LLMs and do not test complex memory hierarchies. To bridge this gap, we propose StructMemEval - a benchmark that tests the agent's ability to organize its long-term memory, not just factual recall. We gather a suite of tasks that humans solve by organizing their knowledge in a specific structure: transaction ledgers, to-do lists, trees and others. Our initial experiments show that simple retrieval-augmented LLMs struggle with these tasks, whereas memory agents can reliably solve them if prompted how to organize their memory. However, we also find that modern LLMs do not always recognize the memory structure when not prompted to do so. This highlights an important direction for future improvements in both LLM training and memory frameworks.