Papers

Knowledge that a language model appears to forget during finetuning often remains stored and can be recovered, a phenomenon called spurious forgetting. Finetuning on new facts can even produce forgetting that undoes itself: recall of the old facts collapses, recovers as training continues on new facts alone, and only then erodes for good. We seek to understand when such forgetting is not catastrophic. A minimal associative memory reproduces these dynamics with three ingredients: keys with shared structure, concentrated new values, and normalization in the network. Finetuning moves all old representations along a common direction, hiding the old facts while preserving their relative geometry; normalization withdraws this shift once the new facts are learned, whereas fact-specific changes accumulate and cause the erosion. Moreover, subtracting the common shift eliminates the collapse in a Transformer trained on synthetic data, and removing a single direction from each weight update restores old facts in a pretrained language model. Forgetting thus combines a shared, reversible loss of access with a slow erosion of individual facts, and only the second is catastrophic. Which one dominates depends on whether the new data move old memories together or apart.

Generative retrieval trains a language model to generate the identifier of a relevant document. Recent work replaces the autoregressive decoder with diffusion, but changes identifiers, training recipe and decoding at once, so differences cannot be credited to the paradigm. On NQ320K and MS300K, we train autoregressive, masked-diffusion and block-diffusion models with residual-quantised, product-quantised and random identifiers. With identifier length and training budget fixed, we decode each model in several ways. Decoding alone moves a diffusion model's Hit@1 by 6.6 to 13.7 points. Our reference diffusion decoding, generate-and-match, generates an identifier, then retrieves the closest corpus identifiers. The generated identifier is right for 14-21% of NQ320K queries. We test one-pass scoring to decode diffusion retrievers: the model reads a fully masked identifier once, and each document is scored by its codes' probabilities. It matches or beats generate-and-match in 11 of 12 settings. Autoregressive models still lead in Hit@1; on NQ320K, the lead comes from the model, not beam search. Starting from one sampled identifier, one-pass scoring removes 46-83% of masked diffusion's deficit to beam search; from generate-and-match, at most a quarter. On NQ320K, every paradigm largely memorises which identifier answers which query: random identifiers keep 83-90% of the Hit@1 of residual-quantised ones. There, product-quantised identifiers lead residual-quantised ones by 3.4 points in the autoregressive model and by -0.7 to +3.6 in diffusion models; across decodings, AR's gap exceeds diffusion's by 1.5-2.3 points, around our 2-point threshold. Paradigm comparisons must report each paradigm at its own recipe and best decoding.

In K-12 mathematics tutoring, student-tutor dialogue provides rich evidence of learners' problem-solving processes and sources of difficulty. Learning analytics research increasingly relies on large language models (LLMs) to extract such information from dialogue for a variety of downstream tasks, including knowledge tracing, behavioral modeling, and diagnosis of student reasoning errors. However, the validity of these model-generated interpretations remains insufficiently understood. In this exploratory study, we examine the validity of LLM classifications of five student failure modes in mathematics tutoring dialogue using an operational diagnostic codebook: uncertainty, misattribution, operator selection, conceptual gap, and procedural slip. Across models, human-LLM agreement was moderate (kappa = .524-.597), while cross-model agreement was substantially higher (kappa = .755-.781; alpha = .769). These findings show that cross-model agreement can create a misleading appearance of correctness, challenging the assumption that consensus among LLMs constitutes evidence of valid learner interpretation. For learning analytics, the implication is clear: scalable labeling is useful only if the inferred constructs are valid, and model consensus cannot substitute for independent evidence of that validity.

Endpoint accuracy on abstract-reasoning benchmarks does not reveal whether a language model has acquired a transferable rule or fit distribution-specific regularities. We study this distinction in small language models on the ARC-TGI benchmark, which organizes abstract grid transformations into controllable task families and supports resampling, spatial shifts, and cross-benchmark transfer. Across more than 1,000 runs, we profile decoder-only, encoder–decoder, and mixture-of-experts model families under supervised fine-tuning. We examine the efficiency and stability of skill acquisition, robustness beyond the training distribution, interactions with model family and task formulation, and layer-wise attention signatures that accompany behavioral differences. Substantial in-distribution accuracy is attainable, but acquisition is sensitive to optimization and unevenly distributed across task families. Performance deteriorates sharply outside the training distribution, including when the rule is retained but grid scale changes. Greater training-set depth and breadth yield uneven gains, while the effect of additional in-context examples depends on model family. Executable-rule induction also yields correct solutions not observed under direct grid generation. On selected tasks, attention diagnostics show distinct concentration and context-dependence profiles, but do not establish general causal mechanisms. Overall, abstract-reasoning scores are conditional on the model, adaptation regime, evaluation distribution, and response format.

Financial reports repeat values across periods, metrics and accounting lines, allowing an LLM-generated calculation to be numerically correct while citing the wrong financial role. We evaluate what probabilistic evidence verification adds beyond number matching using Jev as a source-support verifier for GPT-4.1-mini calculation traces. A signed-number-at-pointer baseline explains most recovery over exact quotation checks. To isolate the remaining role-recognition problem, we hold operands and arithmetic fixed, move citations between same-number cells, and retain controls that express equivalent facts. These contrasts reveal both wrong-role citations that pass and valid alternative citations that are withheld. Explicit column labels improve selected wrong-role decisions while also lowering support for some equivalent evidence. A constructed follow-up on 36 new source pages, labeled by a non-author reviewer, extends this evaluation and exposes the same tradeoff between detecting role errors and retaining valid citations. The contribution is a controlled evaluation that identifies what a probabilistic financial verifier distinguishes when numerical matching is held fixed. For LLM-based financial assistants, it makes numerical correctness, cited-role support and acceptance outcomes separately assessable.

Language models increasingly act as agents. An agent that says an action is wrong and then takes it anyway is a different failure from one that does not know better, and evaluations of stated values cannot see it. We build a pre-registered panel of 248 scenarios across five kinds of pressure. Each scenario is posed twice to the same model, once as the agent choosing what to do and once in the third person asking which option is right, so the model's own judgment is the reference. Every scenario has a twin with the pressure removed, and every model gets a positive control in which its operator orders the violating action, so that a missing gap can be told apart from a blind instrument. On OLMo-3-7B-Instruct, the model takes the action it judged wrong on about one in five pressuring scenarios, more often than on the same scenarios with the pressure removed. Across four instruct models the gap depends on the post-training recipe: OLMo-3 and Meta's Llama-3.1-8B-Instruct carry it; Tulu 3 shows none on the whole panel (above about 0.01 in probability) or on its own most-pressuring scenarios; Qwen2.5-7B-Instruct shows none on the whole panel (above about 0.02) and is unresolved on its own (0.083, -0.028 to 0.195). Meta's recipe and Ai2's Tulu 3 start from the same Llama-3.1 weights, and only Meta's carries the gap. Reading a chat model outside its chat template reverses the sign of its gap with nothing at stake (-0.038 against +0.055 under the template on OLMo-3), a distortion present on two of three recipes. On both models that carry it, reasoning about the stakes before acting moves the choice back toward the model's own judgment, against a same-length non-moral task, with or without the pressure; on OLMo-3, naming the norm at stake does about a third of that. The gap is a measurable target for post-training recipes, not a fixed property of pretrained weights.

Background: Biomedical AI can generate plausible explanations without reliably verifying whether each statement is supported by patient-specific evidence. We developed a neuro-semantic verification framework that converts radiomic measurements into addressable evidence records and machine-checkable claims. Methods: UPenn-GBM radiomics were aligned with de novo CaPTk extraction from standardized MRI and expert-validated segmentations in an independent multicenter cohort. The shared space comprised 1,728 features from T1, T1GD, T2, and FLAIR MRI across three tumor regions. Reference-defined semantic states were derived from 611 UPenn cases. We evaluated cross-cohort transportability, model-linked provenance, deterministic verification, controlled predictive degradation, and an LLM claim-extraction pilot; MGMT prediction served only as a transport stress test. Results: Median semantic-state agreement was 0.786 (weighted kappa 0.709), ranging from 0.918 for morphologic to 0.252 for intensity features. The external evidence ledger contained 1,655 model-linked records for 331 patients. The verifier achieved 100% exact-set accuracy in a 6,620-claim corruption benchmark. In a 24-case pilot, GPT-5.6 Sol reproduced 72/72 prespecified atomic claims, and the frozen verifier recovered 24/24 expected conditions. During controlled degradation, ROC AUC declined from 0.899 to 0.500 while verification accuracy remained 1.000. External MGMT discrimination was weak (ROC AUC 0.543). Conclusions: Verifiability can be engineered and evaluated independently of predictive performance. LLMs may structure explanations, while final evidence-consistency checking remains deterministic.

LLM-based agents solve complex multi-step tasks, but sequential execution incurs substantial latency. In principle, parallelizing work across multiple agents should yield near-linear speedups. Yet existing parallel multi-agent systems often run slower than a single-agent baseline. We attribute this gap to two hidden costs that parallel execution incurs but a serial agent avoids. First, there is a re-exploration cost: redundant effort spent by parallel workers reconstructing context that the orchestrator already possesses, such as prior decisions, that would otherwise be inherited implicitly in a serial execution. Second, there is an alignment cost: the overhead required to reconcile inconsistencies across independently generated outputs. We thus derive a principled decision criterion: a layer should be parallelized only when its critical-path cost, plus re-exploration and alignment overheads, is lower than the corresponding serial cost. While this criterion is naturally expressed in wall-clock time, we observe that LLMs are poorly calibrated when asked to estimate task duration. To address this, we instead measure cost in predicted output tokens, which we empirically find LLMs can estimate substantially more reliably than wall-clock time. Building on this token-based criterion, we propose SquidAgent. It estimates all token budgets in a single planning step, forks each worker directly from the orchestrator's session to eliminate re-exploration cost, and replaces post-hoc reconciliation with a pre-generated shared convention block that converts alignment into a bounded upfront cost. A deterministic scheduler then applies the criterion layer by layer. Empirically, SquidAgent achieves a 2.2$\times$ mean throughput improvement and a 2.6$\times$ mean wall-time speedup over Claude Code, and a 2.0$\times$ throughput improvement over the strongest multi-agent baseline.

Recently Large Language Models (LLMs) and LLM-based agents increasingly need to incorporate knowledge acquired after pretraining, e.g., domain facts, user preferences, documents, and interaction experience. In-context learning (ICL) and ICL-based agent harness remain flexible, but they consume context capacity and incur repeated discretized encoding cost that grows with context length. In-parameter memory offers a complementary substrate: reusable memory information is represented in model parameters, adapters, or other parameter-like objects that are composed into the forward pass at inference time. This survey focuses on methods that augment LLMs with such parametric memory at deployment: a memory-bearing parameter object is plugged into the forward pass during inference, whether it is acquired before or during deployment. We organize the landscape with two orthogonal axes: Parameter Placement, which includes Embedding, Attention, FFN layers, or Hybrid when two or more layers are used; and Parameter Acquisition Time, which distinguishes methods whose memory object is acquired during deployment (online) from those acquired before it (offline). We clarify boundaries, conduct comparisons, and discuss open directions in interference, safety, co-design with ICL, and recursive self-improvement.

Automatic Speech Recognition (ASR) systems often show uneven performance across demographic groups, and errors can be especially difficult to address for speakers belonging to multiple demographic groups. This work studies demographic-aware model merging for fair Speech-LLM-based ASR. Starting from a SLAM-ASR-based model, we fine-tune only the connector on demographic-specific subsets and merge the resulting subgroup-adapted connectors into a global model. We then identify critical cross-axis demographic pairs using subgroup WER and task-vector conflict, and apply intersection-specific correction vectors to the global merged model. Experiments on Fair-Speech show that global demographic merging improves overall WER over the base model, while intersection correction provides additional gains for several merging strategies. In particular, TIES with WER-based correction achieves the best overall WER, reducing it from 7.38% to 5.13%. Subgroup and disparity analyses further show that the proposed approach improves performance across demographic axes, while highlighting that lower average WER does not always imply reduced subgroup disparity.

Emergency messaging such as extreme-weather reports and earthquake instructions can involve high stakes, to the extent that translation errors can lead to tragic consequences. The use of machine translation or generative artificial intelligence might therefore not be recommended. On the other hand, time savings in the initial translation can allow greater investments of resources in revision and authorization processes, as well as a wider range of target languages. An experiment with generative AI translations of an earthquake instruction text from English into Chinese and Spanish shows that use of discourse-specific prompts can considerably improve understandability and actionability, although the translations may still not be trusted by translators. Human revision is still required, not only to detect errors but also because of the ethical need for someone to take responsibility for any errors or delays in such messaging.

Large language models (LLMs) are increasingly relied upon to support ambient documentation and clinical reasoning. Here we examine the impact of a failure mode shared between these two applications by assessing their sensitivity to information incidental to the patient encounter. In 576 patient-clinician dialogues, we found that frontier models inserted small-talk exchanges into 35% of notes, while mean quality scores changed by at most 0.20 points on five-point scales. In 3.7% of frontier notes, models misattributed the asides or used them clinically. In 57 mock recorded consultations, background speech from a separate patient encounter at -10 dB leaked into 48.2% of transcripts, with contamination detected in 5.3% of downstream notes generated by four open-weight models. We propose a dual encoding hypothesis of clinical reasoning and distraction in LLMs, with preliminary evidence that LLM components associated with disruption by incidental information also support clinical reasoning. These findings support evaluating resistance to incidental information before clinical use, with safeguards that prevent contamination while preserving clinical reasoning.

Streaming video-language models must decide not only what to answer, but whether the evidence needed for the current question has arrived. Existing systems learn that decision as a separate trigger; we ask whether an unmodified model already computes it. We show that frozen VideoLLMs carry a linearly readable evidence-readiness signal, labelled from timestamped evidence rather than from model output. It decodes in all seven models of a shared byte-identical evaluation (AUROC 0.733-0.905 under the strictest not-ready sampling, where a fitted clock is near chance), and a probe fitted without any of a benchmark family's footage still reads that family. It is question-conditioned: on byte-identical windows, changing only the question reverses the readout on 66.1% of pairs, while every question-blind control is at chance by construction. The model can answer incorrectly and still encode readiness: AUROC remains 0.722 among wrong answers. Readiness also beats uncertainty estimators and their supervised combination on latency-matched answer selection, and tracks independent human judgments more closely than confidence. Released streaming triggers are also linear readouts, yet a trained trigger read on its own base model's activations is approximately orthogonal to readiness and decodes it far less accurately than a probe. We turn the readout into Readiness Gating, an answer-timing policy that improves accuracy by up to +9.75 pp at matched video duration with negligible computational overhead. How much it gains varies with the accuracy headroom the task makes available: across 26 configurations the gain tracks that headroom, and an intervention that moves it over identical pixels moves the gain with it.

Activation steering has gained popularity as a lightweight inference-time debiasing technique for large language models. However, prior work reports that steering vectors generalise poorly, with unintended effects on model performance and limited transfer to new datasets. Our work analyses what the debiasing direction used for activation steering actually encodes, in order to shed light on its inconsistent performance. We study the linear debiasing direction obtained by contrasting the activations of anti-biased and biased prompts, and evaluate it as a steering intervention across bias and general knowledge benchmarks. We find that this direction is dominated by model confidence, pointing from regions of high to low-probability tokens in activation space rather than encoding a meaningful representation of model bias. Steering along it does reduce measured bias, but this is a consequence of reducing model confidence: on QA benchmarks we find that this steering drives the model to abstain from answering, with a side effect of improving fairness metrics. Our experiments show that model confidence is the dominant separating factor between biased and anti-biased prompts in hidden space, indicating that isolating a linear representation of bias which is disentangled from model confidence is difficult and steering-based debiasing results should be interpreted with care. In short, steering appears to reduce bias, not by correcting the model's underlying preferences, but by making it less confident, even on tasks unrelated to bias.

Sequential test-time training adapts a memory network through successive updates, each computing an inner-loop gradient based on the network's previous state. Intuitively, this state dependence should allow each update to account for what the memory has already learned and better incorporate new information. However, we find that this expected advantage does not consistently materialize in nonlinear memories: a fixed-base parallel TTT baseline outperforms its serial counterpart. Our exploratory experiments point to a key underlying difficulty: nonlinear memories can be harder to optimize than linear ones within a single pass over the sequence. To alleviate this optimization difficulty, we introduce DeltaTTT, which replaces joint inner-loop optimization of a two-layer memory network with layerwise learning. Each layer is assigned a local prediction target and updated through a state-dependent delta rule. This formulation retains a nonlinear readout while enabling chunkwise parallel computation. Experiments on DeltaNet and LaCT backbones show improvements in language modeling and retrieval over their recurrent baselines.

Probes are the workhorse of interpretability. If a model's hidden states predict a variable, the model is said to represent it. But a probe score has no fixed meaning. An $R^2$ of 0.6 may only reflect what the input already gives away, and the same score can mean different things on different data. We propose reading every probe score against two reference points: a floor, what a declared set of simple inputs already predicts, and a ceiling, what the full input can predict. The gap between them, the headroom, is the range in which a probe can show that a model computes something beyond the simple inputs. We prove that headroom vanishes in two ways: the target stops depending on a hidden variable the model must infer, or the input stops revealing it. We test this on transformers trained for in-context meta-analysis, which must infer the hidden heterogeneity between studies to weight them correctly, and where both reference points are known. Under distribution shift, probe scores fall and prediction error rises $12$–$15\times$, yet the model recovers a similar share of the headroom, indicating that the data lost information, not the representation. We then analyze the real models. The single-cell foundation model scGPT encodes biological variability only partially. We also revisit four influential LLM probing studies, which claim that models represent geography, the state of an Othello board, truth, and the demographics of their users. Against a floor computed from the input text alone, some of these claims hold, while others are largely explained by the text itself.

Whether alignment gets easier or harder as models grow is often argued from isolated findings, as if alignment were one property. We treat it as a family of measurable scaling relations: for each risk category r, the alignment burden needed to hold a fixed safety target is modeled as B_r(N)=a_rN^alpha_r, with N a capability proxy; against a budget proportional to N, scaling helps if alpha_r<1, keeps pace if alpha_r~1, and accumulates alignment debt if alpha_r>1. We give three operationalizations of burden and distinguish observed, audited and true alignment. A toy model, in which corrections consume capability headroom, makes the consequences explicit. We prove that the largest exponent among corrected risks, not an average, sets the long-run regime; that above 1 any policy holding headroom above a floor must grow super-exponentially; that, for burdens that are positive mixtures of power laws, fits on small models underestimate large-scale exponents; and that an audit that uncovers hidden failures without false positives never underestimates true alignment. We propose a pre-registrable protocol and apply reduced versions of it twice. A preregistered reanalysis of public adversarial-training data for Pythia classifiers finds that the compute needed to bring attack success under 10% grows as N^0.60. A preregistered pilot on Qwen2.5 0.5B-72B finds exponents of -0.05 for truthfulness and 0.48 for stated dispositions (both scaling helps under its reduced rule, though local slopes approach 1 at the top; replicated on Qwen3 0.6B-14B), while sycophancy (0.89, or 0.83 with two seeds added at 72B) and a planted backdoor are undetermined: the backdoor is removed quickly when its trigger is known but survives blind safety training at four of five sizes. We release four browser games that play these laws (www.aisafety.fun). We make no claim about which regime holds for current frontier models.

LLMs are increasingly deployed as autonomous agents in social environments, making it critical to study their ability to faithfully simulate human interactions. Central to this is grounding agents in realistic user personas, yet existing datasets rely on fictional personas and are limited to a handful of languages, lacking the empirical grounding necessary to evaluate behavioral fidelity across diverse populations. We introduce Wiki-Talkie, a multilingual dataset of real-world conversations from Wikipedia Talk pages across five languages spanning two language families: Germanic (German, English) and Romance (Spanish, French, Italian), paired with personas derived from real user communities and encompassing sociodemographic attributes, self-descriptions, and behaviorally grounded interaction traits. Using Wiki-Talkie, we evaluate agent interactional behavior on a next-turn generation task across various persona conditioning strategies. Our evaluation assesses whether agents collectively reproduce the distributional behavioral patterns observed in human discussions. Results show that user's comment history exemplifying interaction behavior consistently outperforms explicit persona information. In addition, models systematically underproduce negative or extreme sentiments, while over producing references and suggestions, revealing biases toward agreeableness and positivity. Crucially, these patterns hold robustly across languages, with small cross-lingual differences.

Depression has no diagnostic blood test. Language models promise tireless, consistent assessment, but can accurate raters disagree about individuals? We pre-registered 880 language-model raters, crossing 11 open models with prompting and scoring choices, and applied them to 189 interviews against the eight-item Patient Health Questionnaire. Model choice explained 30.0% of summed-symptom score variance, stable participant differences 10.5%. Two randomly drawn raters with area under the receiver operating characteristic curve (AUC) >= 0.70 disagreed on screening decisions for 40% of participants, on average. Average over-rating governed how many were flagged, yet equal-capacity raters chose differently for about one participant in five. A locked analysis of 86 new interviews reproduced the main pre-registered findings. Exploratory recalibration with 40 labelled participants raised accuracy from about 60% to 75% and halved disagreement, leaving one participant in five decided differently. Calibration repaired much of the rater dependence without securing agreement about individuals.

Long-context inference and Retrieval-Augmented Generation (RAG) handle evidence selection at vastly different scales, from a single long prompt to an entire corpus. We ask whether a single model-internal mechanism can select evidence across this range. We introduce UNifying REtrieval And Long-Context with a Single Model (UNREAL), a model-native evidence selection framework to span corpus retrieval and long-context inference. UNREAL encodes chunks and derives retrieval queries directly from the frozen LLM's internal representations. It adds fewer than 500K trainable parameters and leaves the backbone unchanged. On a 3B-token, 21M-chunk Wikipedia index, all four dense and hybrid UNREAL backbones outperform state-of-the-art retriever-reranker systems. The best model raises recall from 49.1% to 73.2% on HotpotQA, from 31.7% to 60.1% on 2WikiMultiHopQA, and from 8.8% to 14.4% on MuSiQue. Applied to long-context tasks, the same selection mechanism removes distractors before generation, raising NoLiMa accuracy from 1.0% to 24.83% at its maximum context length of 128K tokens, and LV-Eval's F1 score from 49.97% to 54.66% at 256K. UNREAL also reduces FLOPs and time-to-first-token relative to full-context inference from roughly 32K tokens onward, with larger gains as context grows. Together, these results establish model-internal evidence selection as a common foundation for corpus retrieval and evidence-sparse long-context inference.

Retrieval-augmented generation (RAG) is a widely used approach for grounding large language models (LLMs) in external knowledge. However, configuring a pipeline is an expensive hyperparameter optimization problem over many interacting choices, from chunking and embedding model to reranking and generation. Existing optimizers, from greedy search to Bayesian optimization, reduce each trial to an aggregate score and search without modeling why a configuration performed as it did, even though the retrieved chunks already provide evidence about whether each failure occurred during retrieval or after it. We introduce Agentic AutoRAG, an LLM-agent optimizer for multi-objective RAG hyperparameter optimization with retrieval-versus-generation failure attribution. It proposes configurations scored on a frozen exam from the corpus: after each trial a Diagnoser attributes each failed question to retrieval or generation, and a Proposer, grounded in a knowledge base of model rankings and pricing, selects the next configuration, weighing accuracy against cost to trace a Pareto frontier. On three multi-hop QA benchmarks it reaches higher LLM-judge accuracy than every baseline we compare, and within its first 10 trials it matches or beats the statistical baselines' full 30-trial judge accuracy. In its cost-aware mode on a real-world healthcare corpus it reaches a median exam accuracy of 77%, above the strongest baseline's 71.5%, at about 58% of that baseline's cost per query, and it matches that 71.5% at about 22% of the cost.

On-Policy Distillation (OPD) trains a student on its own generations using teacher feedback. With different tokenizers, comparing teacher and student predictions requires alignment at both sequence and vocabulary levels. In this paper, we examine whether expanding this alignment coverage improves learning. Across three heterogeneous teacher–student pairs on mathematical reasoning and code generation, strict 1:1 groups already cover most student-generated tokens despite substantial vocabulary mismatch. On responses sampled from the students before distillation, the shared vocabulary retains nearly all teacher and student probability mass at strictly aligned positions on average. Restricting reverse KL to a student-selected top-16 subset of the shared vocabulary at each strict position achieves accuracy comparable to full shared-vocabulary OPD, outperforming the evaluated cross-tokenizer baselines. Adding mean squared error supervision on span log-probabilities in mismatch groups gives complete supervision coverage, yet reduces accuracy. At checkpoints from training with only the strict loss, the span gradients show weak or negative directional agreement with the strict gradients and grow in magnitude relative to them. These diagnostics may help explain the accuracy drop from adding span supervision. Our findings motivate a shift from maximizing alignment coverage to prioritizing supervision reliability: compact supervision at strict positions can be more effective than broader coverage that introduces weakly aligned or conflicting training signals.

Large language models routinely answer questions that cannot be answered from the information given, and in dialogue they answer before enough has been said. Unanswerability is linearly decodable from hidden states, but it is unclear which of its forms share a representation and whether the signal is useful in dialogue. We contribute a turn-labeled multi-turn benchmark (423 conversations, 1,661 labeled turn-states) and an evaluation harness with a simulated user who answers clarifying questions, and use them with six datasets and six open-weight LLMs to test how far probes for unanswerability carry. Probes transfer robustly between datasets that share a ground of unanswerability: missing information in math (AUROC 0.77-0.97) and in a passage (SQuAD 2.0<->MuSiQue, 0.77-0.90). Probes for epistemic "known-unknowns" transfer poorly to math, but this separation weakens under lexical controls and changes with layer and coordinate system, so it remains unresolved. Single-turn probes fail zero-shot to detect when a conversation becomes answerable; in-structure probes recover it, but no better than a bag-of-words classifier. A gate on the calibrated probe, with no model fine-tuning, fires on underspecified turns far more precisely than chance, and its end-task success comes within 0.08 of a gate given the true labels. Yet across four models it does not reliably beat vanilla generation or prompted consolidation. The remaining gap lies mostly in how models use a clarification, not in detection.

Large language models (LLMs) have demonstrated strong capabilities in question answering, yet they still frequently suffer from hallucinations on knowledge-intensive tasks. Knowledge graphs (KGs) provide LLMs with structured, interpretable, and updatable factual grounding, making them a promising external knowledge source for reliable reasoning. However, existing LLM-guided graph reasoning methods typically rely on hop-wise greedy or beam-style pruning during evidence retrieval. Such local decision processes are inherently myopic: evidence that appears weak near the source may become crucial only after deeper graph context is explored, causing answer-critical branches to be discarded prematurely and making the reasoning chain difficult to recover. To address this limitation, we propose Foresight-over-Graph (FoG), a foresight-aware evidence retrieval framework for knowledge base question answering (KBQA). FoG iteratively constructs a question-relevant evidence subgraph and uses far-to-near feedback to guide path exploration, and maintains a compact memory subgraph to support continued exploration. Extensive experiments on widely used KBQA benchmarks demonstrate that FoG achieves state-of-the-art performance, with a particularly large improvement of 16.58% in Hit on CWQ, while also reducing LLM calls and token usage. Our code is available at https://github.com/yhong7/FoG .

Continual learning (CL) is essential for Large Language Models (LLMs) to sequentially adapt to evolving tasks. To mitigate catastrophic forgetting, recent advances implement low-rank adaptation with orthogonal projections (e.g., O-LoRA) to isolate task parameters. However, we reveal that such strict geometric constraints trigger an "Orthogonality Dilemma": rigid parameter isolation impedes the transfer and accumulation of shared representations across semantically related tasks. In this work, we propose a new replay-free method, called Consolidation and Decoupling LoRA (CoDe-LoRA), for CL of LLMs. CoDe-LoRA disentangles the learning process into Consolidating Universal Knowledge and Decoupling Task-Specific Knowledge. To achieve this, CoDe-LoRA leverages an adaptive null space projection mechanism and semantic routing to balance knowledge accumulation with task-specific adaptation. Experimental results across four backbones and three CL benchmarks show that CoDe-LoRA achieves the best average accuracy. Our code is available at https://github.com/Estrellajer/CoDe-LoRA.

Current evaluation of multilingual Large Language Models (LLMs) rests on an implicit Translation-Isomorphism Assumption (TIA): that semantic structures across languages are congruent and mutually mappable without loss of information. We argue that this assumption is not merely violated in practice, but ill-posed in principle for a typologically identifiable class of concepts, including pragmatic markers, honorifics, and diachronically stratified terms. We formalize this failure using a usage-cloud framework, representing concepts as point sets of contextualized embeddings. We define $\alpha$-unalignability as the impossibility of any mapping that simultaneously preserves lexical faithfulness (centroid correspondence) and structural faithfulness (local neighborhood topology). We provide three layers of evidence. Behaviorally, we show that FLORES-200 translation failures are predicted by language family and resource class but not by script, and that LOBSTER reasoning scores vary by family. Mechanistically, we report a Representation-Intervention Gap (RIG) in a nine-model case study on Yami: the models' activations encode a regularity along which Yami groups with other low-resource and Austronesian languages, yet interventions on language-specific neurons show no demonstrated advantage over random masks: the regularity is visible but not usable by this intervention. Finally, we operationalize these findings into a multidimensional diagnostic profile: Cycle-Consistency, Pragmatic-Load Disagreement, Manifold-Curvature Mismatch, and RIG. We argue that collapsing cultural competence into a single scalar incentivizes "probabilistic flattening," and that recognizing the unalignable class is a precondition for AI that respects, rather than erases, cultural divergence. This suggests that multilingual alignment is not a single well-defined objective, but a set of mutually incompatible projections.

Long-term conversational memory is becoming an integral component of modern LLM systems. Proposed architectures group records by topics and events, construct hierarchies and graphs, and connect facts through causal and temporal relations. We experimentally study the interaction between two memory parameters: structural depth and the width of context supplied to the answer model. Using EverMemBench, we evaluate depths D1-D4, core budgets of 1,024/2,048/4,096 tokens, and additional Production and Oracle conditions up to the full archive. Increasing width from 1K to 4K improves Accuracy by 10.11-17.98 percentage points, whereas increasing depth provides no monotonic gain. Beyond 8-16K, Production performance reaches a plateau while tokens per correct answer continue to increase; Oracle preserves quality on full archives of 68-71K tokens. These results motivate further investigation of large, coherent context blocks instead of progressively deeper memory structures.

We introduce STRUCTURALCOST, a self-paced reading dataset of 475 participants and 40,800 observations isolating the processing cost of long-distance subject-verb dependency resolution. We replicate a low-powered psycholinguistic finding at NLP scale, namely that human reading times at the main verb increase with dependency length, driven by syntactic embedding beyond linear distance. Different language models – spanning n-gram models, SSMs, and transformers – partially mirror this graded difficulty profile, yet underestimate the integration cost humans incur, with a gap that persists across architectures and model sizes. This suggests these models capture the predictive component of human processing but not the full integration cost that working memory imposes. STRUCTURALCOST provides data needed to drive progress toward evaluating the cognitive plausibility of language models.

Low-rank quantization error compensation (LQEC) recovers the accuracy lost under aggressive weight quantization by attaching a closed-form rank-$r$ adapter beside each frozen quantized weight, without any training. We show that existing compensators are limited by two shared simplifications. They calibrate symmetrically, evaluating the full-precision and compensated weights on the same activation, which yields a compensation target that is inherently high-rank – so a fixed rank budget captures only a small fraction of it. And they minimize only the second-order term of the loss, although the compensated model is not stationary: a first-order descent direction larger than the applied compensation itself remains in every layer, and no reconstruction objective can absorb it. We propose a two-stage closed-form framework that removes both simplifications. Stage 1 aligns each layer's output with the full-precision model under a Fisher-weighted asymmetric objective, concentrating the rank budget on a rank-compressible target. Stage 2 re-measures statistics on the compensated model and applies a rank-constrained natural-gradient step that absorbs the remaining first-order signal. Every adapter is the result of a single truncated SVD; backward passes serve only to collect statistics. At 2 bits under QuIP#, our method reduces WikiText-2 perplexity from 12.43 to 10.26 on Qwen3-8B and from 21.11 to 13.22 on Qwen3-4B. On the held-out C4 corpus, it recovers 51% and 84% of the gap to FP16, versus 31% and 63% for the strongest baseline, with consistent gains in the seven-task zero-shot average, at higher bit-widths, and under a distinct quantizer.

Fine-grained emotion recognition supports therapy tools and social robots, but it needs facial data, which raises privacy and data-protection concerns. EmoNet-Face-HQ answers that with generated portraits, expert-rated over a $40$-category taxonomy far finer than the usual six to eight basic emotions. Under the protocol it ships with, vision-language models (VLMs) score poorly on that taxonomy, and the benchmark concludes that a dedicated fine-tuned model is necessary: Empathic-Insight-Face (EIF; Small/Large). We show that off-the-shelf VLMs match or beat that fine-tuned model when the answer is not generated but read from the logits, as one binary query per category. We keep the benchmark's images, taxonomy and ratings, and change only how the answer is read. Experts agree at $\kappa_w = 0.468$ on the five categories they measure most reliably. Generatively, no interval among eleven open-weight VLMs lies entirely above that anchor ($\kappa_w=0.268$-$0.486$). Under verification all eleven clear it, each of them significantly better at $\kappa_w=0.507$-$0.586$. Three also significantly beat EIF sitting at $\kappa_w = 0.551$ (Small; $0.534$ Large). The gain comes from the graded probability and not from asking a yes/no question: as a control, thresholding those same probabilities to yes/no costs 142% of the average gains and drops binarization below generative elicitation to $\kappa_w=0.254$-$0.423$. A replication on real photographs (FACES) is weaker and mixed: of the ten models that pass a validity gate, six gain, three are neutral to positive and one is negative, so the effect is not confined to synthetic data.