LLM 理论进展

Large language models (LLMs) can produce factually incorrect answers with high confidence, undermining their reliability and limiting the effectiveness of uncertainty-based error detection. While prior research attributes confident hallucinations to factors such as missing knowledge in training data, reasoning errors, or stochastic decoding, we uncover that post-training alignment itself is a primary driver of these errors, a phenomenon we call the Alignment Paradox. Across five model families evaluated on factual benchmarks, unaligned base models produce few high-confidence errors on long-tail factual queries, whereas instruction-tuned models multiply high-confidence errors ($p \ge 0.95$) by more than an order of magnitude (10$\times$ to 35$\times$). Layer-wise probing with the Logit Lens reveals that this overconfidence emerges in late layers, where wrong-answer margins expand past 4.0 points after remaining near zero across early and intermediate layers. These findings motivate limiting margin growth during post-training. We implement this principle through an entropy-dependent margin bound in direct preference optimization (DPO). In multi-epoch experiments with Mistral-7B, the bounded objective reduces high-confidence errors by up to 35.3% relative to standard DPO while maintaining performance on evaluated general reasoning benchmarks. These results show that bounded margins mitigate confident hallucinations during post-training.

Large language models (LLMs) can solve every intermediate step of a multi-step math problem on its own and still fail the full problem, even when given a roadmap of the steps and all of their answers. We introduce OracleLadder, a diagnostic evaluation that locates where LLM math reasoning fails by giving the model increasing levels of oracle help. For each problem, a teacher model writes a fixed roadmap of intermediate sub-goals (milestones), and a deterministic symbolic verifier grades every answer. Testing the model with no help, with the roadmap, with the roadmap plus the milestone answers, and on each milestone alone sorts each failure into one of five reasoning gaps. On 354 NuminaMath problems and six models from 8B to 671B parameters (Qwen3, gpt-oss, Llama 3.3, DeepSeek-V3.1), the largest gap for every model is the composition gap, a stricter form of the compositionality gap. It covers 33-48% of problems, and 24-37% after removing problems that an LLM review flags as grading errors. Accuracy and milestone-help recovery rank the two strongest models differently, and two RLVR runs with similar accuracy gains move problems differently. The roadmap effect replicates on MATH500 and AIME 2024/25, per-problem recovery agrees for 83-87% of problems under an independent second teacher, and the help ladder carries over to code generation. We release the data, roadmaps, prompts, and code at https://github.com/slark-prime/OracleLadder.

Large Language Model (LLM) alignment is intended to ensure that models remain helpful and safe, but its stability under input distributional shift is not yet fully understood. Prior work shows that aligned models can fail under jailbreak prompts, alternative encodings, and cross-lingual transfer, yet these failures are usually studied as attacks rather than controlled probes of alignment generalization. Moreover, existing evidence is largely grounded in natural language variation already represented during pretraining, leaving unresolved whether alignment generalizes with semantic content or remains tied to superficial surface patterns. In this paper, we study this question using synthetic semantic-preserving transformations that are rule-based and invertible, preserving task-relevant meaning while shifting inputs beyond standard linguistic variation. Across four open-weight and four commercial models, under both fine-tuning and in-context learning, we use these transformations as a probe of alignment generalization and identify an empirical pattern we term Alignment--utility asymmetry: once models can operate effectively on transformed inputs, task utility is often substantially retained while alignment failure increases more sharply. For example, adapted GPT-4.1 mini shows only limited utility degradation under transformation while its harmful rate rises from 13.3 to 74.3; Gemini 3 Flash similarly retains near-original utility while its harmful rate increases from 2.3 to 43.0. Taken together, these results suggest that semantic-preserving distribution shifts can expose a recurring gap in how utility and alignment generalize in current LLMs.

Vision-language models (VLMs) have demonstrated excellent performance in visual tasks, but their visual causal reasoning capabilities still lack reliable evaluation. Existing evaluations struggle to distinguish whether a model is performing causal reasoning based on visual evidence or relying on statistical correlations for shortcut learning, thereby potentially overestimating their actual capabilities. This paper proposes CCRV-Bench, a constraint-driven visual causal reasoning benchmark for single-image physical scenarios. We construct an orthogonal framework that evaluates four causal task dimensions: causal relation discovery, state prediction, causal diagnosis, and intervention. We further introduce entity symbolization, spatial grounding, the factual adversarial constraint, and minimalist output constraints to reduce shortcut cues while preserving the physical commonsense required by the task. Experiments across 15 multimodal models show that constraint sensitivity is task- and model-dependent: intervention has the largest average effective degradation among the four causal tasks, spatial grounding is the most damaging constraint on average, and the factual adversarial constraint improves DCR for all evaluated models. These results show that unconstrained performance does not determine constrained robustness and that a single aggregate score can obscure distinct failures in causal identification, spatial grounding, and constraint-compliant expression. CCRV-Bench provides a standardized framework for diagnosing image-grounded causal reasoning under controlled constraints. The code is available at https://github.com/0815linyuan/CCRV-Bench-Constraint-Based-Evaluation-of-Causal-Reasoning-in-Vision-Language-Models

We present FLIP, a final-layer inference-time probe for testing whether a logit-facing intervention site in an open-weight vision-language model (VLM) supports structured, task-linked computation rather than generic perturbation. Behavioral change under internal intervention is otherwise mechanistically ambiguous: it may reflect improved use of visual evidence, generic output instability, or outright degradation. FLIP applies elementwise flooring to the final normalized hidden state before logit computation, leaving parameters, prompts, and decoding unchanged. On a controlled detection/counting probe, sweeping intervention strength reveals three regions: negligible change, a bounded interior regime in which detection recall at IoU 0.50 ($R_{50}$) improves while tolerant counting error ($\mathcal{E}_{\mathrm{count}}$) falls, and over-suppression. We formalize a four-criterion probe-and-sweep protocol for disciplining the interpretation of intervention effects: regime structure, grounding-proxy alignment, feature-coherence dependence, and failure to reproduce the same positive regime on a performance-based negative control. The post-normalization state passed to the output head is the logit-facing instantiation of this test; under a non-targeted flooring sweep it satisfies the full protocol. Raw decoder-layer interventions, including the last-block output before final normalization, and the singleton-pair left/right control fail to reproduce the Final-site signature, while same-site operators and multiple VLMs replicate it. FLIP is therefore a validation step for intervention-based mechanistic interpretability, not a steering method. Repo at https://github.com/earl-juanico/flip-qwen3vl/

The linear representation hypothesis (LRH) has become a standard lens for measuring and intervening on semantic information through the internal representations of large language models (LLMs). A growing body of work has begun extending this perspective to vision-language-action (VLA) models, but the dynamical nature of embodied interaction introduces an additional challenge. Unlike semantic attributes commonly studied in LLMs, such as gender or language, a physical quantity of interest (QoI) in a VLA evolves jointly with the system dynamics: the representation influences the actions selected by the policy, which alter the physical state and, in turn, the next representation. In this paper, we develop a theoretical, signature-based formulation of the LRH for VLA that unifies representations and policies. On the representation side, we establish the existence of representations from which the future evolution of a QoI under a candidate action trajectory can be recovered via linear probing. On the policy side, we introduce a signature generalized linear model for stochastic action chunks. This structure yields a monotonic change in the expected future QoI along linear paths in natural parameter space, enabling linear steering. We first validate this structure in a controlled oracle setting, then examine whether the same probing and steering mechanisms emerge in a pretrained VLA.

We introduce effective depth ($\Deff$), a scalar diagnostic that treats the layer-wise residual stream of a transformer as a discrete-time process, measures how representation similarity decays with layer distance, and aggregates that profile into one number. Across sixteen decoder-only language models, $\Deff$ separates a structural consequence of residual accumulation from an empirical one: even maximally diverse orthogonal updates have the closed-form reference $F_L = 2L/(L+1)<2$, yet fifteen of sixteen default measurements lie below $F_L$ (Qwen3.5: 32--44%, OLMo-2: 40--41%, Pythia: 23--28%). Matched references show that the gap is not caused by the persistent initial state or update-size imbalance, but is largely a calibrated signature of correlated residual updates rather than evidence that depth is unused. Symmetric position-0, token-normalisation, and top-PC controls show the regime is not reducible to BOS or top-PC artefacts: the lone above-reference default outlier joins the same regime, and all sixteen models are sub-reference after token-normalisation or top-1-PC removal. Intermediate checkpoints show that the regime is established early in OLMo-2 and stable through 5T tokens, while Pythia-1.4B follows a distinct decreasing trajectory. A controlled residual-carry intervention supports the mechanism, and $\Deff$ is best read as a global accumulated-state diagnostic, not as a capability score or pruning method.

Chain-of-thought (CoT) monitoring is a promising safety technique for reasoning models, enabling detection of problematic reasoning before models act. A key concern is encoded reasoning, where models hide their true reasoning in ways that monitors and humans cannot interpret. Optimization pressure from CoT monitors during reinforcement learning is considered a likely driver of such behavior. We investigate this by training reasoning models to perform a main task and a side task, while penalizing them when a monitor detects reasoning about the side task. Surprisingly, models learn to evade monitors without encoding their reasoning. Instead, they learn to phrase and format their chains of thought such that monitors fail to flag side task reasoning, while the reasoning remains completely transparent to human readers. We call this phenomenon monitor jailbreaking. We find that monitor jailbreaking arises across different model sizes, monitors, and tasks. Jailbreaks generalize to monitors not seen during training, including both less and more capable monitors, and transfer across different monitor prompts. While jailbreaking strategies appear simple, manually replicating them does not reliably fool monitors. Finally, we show that paraphrasing is an effective defense: paraphrasing a jailbroken CoT allows the same monitor to correctly flag it, while still allowing the model to perform both tasks.

Most Vision-Language Models (VLMs) are built by extending pretrained Large Language Models (LLMs) with visual modules and multimodal alignment. However, this multimodal scaling often degrades the language-side reasoning ability originally encoded in the base LLM. While the base LLM retains usable reasoning after scaling, the aligned VLM itself cannot reliably access this ability. Therefore, recovering the degraded reasoning capability in VLMs would benefit more from seeking help from the base LLM than from the VLM alone. Motivated by this, we propose LIFT (Language-side reasonIng Facilitation and Transfer), a lightweight vector-intervention method that transfers reasoning capability from the base LLM to the VLM without retraining the backbone. LIFT defines Reasoning Vectors as answer-token hidden-state differences between a Reasoner path with an explicit reasoning trace and a Solver path without it, and injects these vectors into language-side activations of the target VLM. LIFT further supports learnable vector adaptation while keeping the VLM backbone frozen. We evaluate LIFT on two VLMs across six reasoning benchmarks, comparing Reasoning Vectors extracted from the base LLM and from the aligned VLM under matched protocols. Results show that LLM-derived vectors consistently outperform VLM-derived vectors, confirming that the base LLM is a more effective source for recovering reasoning. LIFT partially recovers degraded reasoning through lightweight language-side interventions. Further analyses show that Reasoning Vectors influence intermediate reasoning behavior rather than merely altering final answers. The source code will be released soon.

Validating an LLM-as-a-judge requires estimating its agreement with humans, yet annotation budgets rarely allow every item to be multiply labeled. We prove that this overlap sparsity is the first-order determinant of wrong deployment decisions: at 5% pairwise overlap, wrong-decision rates reach 25% and the probability of selecting the wrong best judge among ten candidates is 65%. The two actionable levers are overlap quantity and allocation. For quantity, we derive a minimum-overlap formula showing $ρ\geq 0.25$ suffices for non-borderline judges while borderline cases remain fundamentally hard. For allocation, a zero-cost stratified scheme halves false-rejection rates relative to random sampling when strata are informative. We validate on 10 LLM judges across four evaluation matrices spanning visual assessment, causal reasoning, and summarization.

Medical report generation has made significant progress with the rise of modern vision-language models and the growing availability of large-scale medical datasets. However, hallucinations remain a major challenge, largely due to the limitations of supervised fine-tuning (SFT), which prioritizes lexical similarity to reference reports rather than clinical correctness. While reinforcement learning has shown strong performance in domains with verifiable rewards such as mathematics and code generation, its application to open-ended medical tasks remains limited. Existing work focuses on evaluating final answers, overlooking the model's reasoning, despite evidence that flawed reasoning can degrade overall performance. In this study, we propose a framework that verifies the model's reasoning process by integrating anatomical regions, bounding boxes, and region-level textual descriptions. We design spatial and factual reward mechanisms to ensure that the model's reasoning is both visually grounded and factually accurate. Starting from Qwen3-VL-8B-Instruct as our base model, we adapt it to the medical domain using supervised fine-tuning, introduce reasoning capability through a cold-start SFT stage, and refine it with reinforcement learning. We find that RL provides performance gains beyond those achievable through SFT alone, and that jointly verifying both reasoning steps and final outputs yields larger improvements than verifying either in isolation. We further identify multiple modes of reward hacking in the RL stage. Finally, the model's structured think traces enhance interpretability, making its outputs easier to audit for clinical use.

Reasoning segmentation aims to interpret implicit textual queries and enable fine-grained visual perception, which is critical for applications such as human-computer interaction and embodied agents. Existing methods typically generate explicit Chain-of-Thought (CoT) by multimodal large language models (MLLMs) before localizing the target. Although intuitive, such explicit verbal reasoning introduces substantial attention interference: redundant textual tokens disrupt attention during perception-token generation and also increase the effective distance between visual tokens. To address this issue, we propose LIRSeg, which fully replaces explicit CoT with a compact set of learnable latent tokens for reasoning segmentation. LIRSeg is trained in two stages: spatial alignment grounds the latent tokens in object-relevant visual evidence, and GRPO further optimizes them with segmentation rewards. To make these compact latent tokens more informative, we introduce three complementary mechanisms from an information perspective: extreme-advantage sampling for selecting informative training signals, decoupled exploration-stability updates for learning complementary representations, and latent diversity amplification for preventing representational collapse. Extensive experiments on benchmarks demonstrate that LIRSeg consistently improves both segmentation accuracy and reasoning efficiency. Compared with the VisionReasoner baseline, LIRSeg achieves absolute gIoU improvements of 4.9% on ReasonSeg, 7.1% on MUSE, and 4.7% on MMR, while achieving a approximately 16x reduction in reasoning tokens. Code is available in supplementary materials.

Large language models (LLMs) implicitly infer attributes of their users and adapt their behavior accordingly, yet these beliefs remain difficult to inspect and causally manipulate. We introduce Belief Self-Distillation (BSD), a unified read-write framework that bridges linear and causal probing by learning a compact user representation that can be both decoded and written back into the model. The frozen LLM acts as its own teacher, distilling beliefs from natural conversations without external annotations. Unlike conventional probing, BSD isolates not only information present in activations, but a state whose causal role can be directly tested. Across multiple model families, BSD faithfully recovers user beliefs and enables substantially stronger interventions than matched hidden-state steering. Crucially, we find that refusal depends not only on the request, but on the model's inferred user intent: changing this belief alters refusal while holding the request fixed. We further uncover a striking cross-model regularity: independently trained LLMs converge on a shared geometry for representing their users. Together, these results reveal implicit user models as readable and causally writable internal states with direct implications for AI safety, shaping how models condition safety decisions on whom they believe they are interacting with.

Activation steering is a powerful training-free paradigm for controlling large language models at inference time. However, standard approaches estimate a per-layer steering direction from contrastive data and apply it on the layer's entire representation space, which may couple the intervention to off-target properties present in the contrastive data and degrade unrelated capabilities. To mitigate this issue, we introduce LocUS (Localized Unembedding Steering), a method which grounds activation steering to the model's own output vocabulary subspace. By identifying a property-specific linear subspace within the unembedding matrix, LocUS enforces a geometric constraint that restricts the steering transformation to a specific subspace and at the same time localizes its application to a sparse subset of attention heads. Extensive evaluations across three model families on toxicity mitigation, sentiment redirection and sycophancy suppression show that LocUS matches or outperforms state-of-the-art baselines while intervening on under 6% of parameters and better preserving general capability.

Large language models operating in multilingual contexts must resolve target response languages early in generation, yet the causal circuitry governing first-token language identity decisions remains poorly mapped. We present an end-to-end structural circuit analysis across six model architectures spanning four families: GPT-2, BLOOM-560M, Pythia-1B/2.8B, and Qwen2.5-1.5B Base/Instruct. Using Edge Attribution Patching (EAP) with FP16 active clamping, followed by exact activation patching verification with a 2,000-candidate-edge search ceiling, we extract directed acyclic graphs driving first-token language broadcasting. Across the standalone models, we observe deep or mid-to-deep broadcasting hubs, though the evidence is strongest for Pythia-2.8B and BLOOM-560M because GPT-2 and Pythia-1B leave few out-of-graph heads for comparison, while both Qwen2.5-1.5B variants invert the necessity check. Scaling from Pythia-1B to 2.8B expands node participation while maintaining a similar verified edge budget, producing sparser topology. The Qwen2.5-1.5B base and instruct circuits retain 84.7% Jaccard similarity, including the Layer 27 hub, indicating that first-token routing is largely established during pretraining and preserved by instruction tuning. Finally, EAP scores correlate weakly with exact patching deltas across most models, showing that linear gradient approximations can diverge from causal interventions in FP16 and motivating exact-patching verification for reliable circuit discovery.

The empirical success of pretraining large language models has inspired a deeper investigation into the underlying loss landscapes and the optimization dynamics. Recent empirical and theoretical study suggest that the training loss landscape often exhibits a "river-valley" structure, which features a low-loss manifold (river) flanked by sharp orthogonal directions with higher loss (mountains). In the long term, the optimization progress is determined primarily by the progress along the river. Within such a landscape, gradient descent with large learning rates can move faster along the river despite high apparent loss due to vertical oscillations, while a subsequent sharp decay in the learning rate suppresses these oscillations, revealing genuine optimization progress. This explains the recent success of warmup-stable-decay (WSD) learning rate scheduler which, unlike cosine scheduling, keeps stable high learning rate and decays before producing intermediate checkpoints. Building on this foundation, in this work we take a step further and study the role of momentum within such a loss landscape. We establish theoretical analysis that characterizes how momentum accelerates optimization by stabilizing large learning rates that can not be tolerated by vanilla GD without deviating significantly from the river. The enabled large learning rate in-turn gives greater speed along the river and makes faster essential progress in the long run. Another intriguing observation from theory is that for a river-valley landscape with very flat and slow-spinning river, the momentum itself does not contribute directly to acceleration in terms of the speed of tracking the river, while the main acceleration comes from the admissible larger learning rate.

Corr2Cause asks whether a causal claim holds in every DAG compatible with observed correlations and conditional independencies. We frame this as latent-object reasoning: the label is defined by a CPDAG query, but free-form chain-of-thought often collapses the Markov-equivalence-class problem into local pattern matching. We propose Structured Thinking, a two-turn pipeline that first externalizes a typed, schema-constrained CPDAG summary and then answers against that graph state. On the Corr2Cause full test, Structured Thinking raises Qwen3.5-27B from $73.0$ to $86.4$ $F_1$(Yes) over a strong PC-instruction baseline in the primary paired run ($+13.4$ pp; McNemar $p=2.4\times 10^{-6}$; bootstrap $95%$ CI [$+8.4$, $+18.6$]); across three full-ID seeds, the mean gain is $+8.1 \pm 5.3$ pp. A PC-scaffolded two-turn prose control reaches only $67.6$ $F_1$, indicating that a detailed PC scaffold plus a schema-free prose intermediate is not sufficient. The same pattern holds on Qwen3.6-27B, Paraphrase-OOD, and GPT-5.4-mini. Scrambling the emitted CPDAG costs $12.0$ pp $F_1$, and a full-split audit shows close agreement with the reference CPDAG (ID skeleton $F_1$ $0.960$; exact match $75.9%$). These results support a bounded design principle: externalize the latent object that defines the label, constrain its form, and test whether downstream answers use it.

Collaborative science learning requires nuanced interpretation of student dialogue to characterize how learners identify knowledge gaps, build explanations, and work toward resolution - a theory-driven analysis that is labor-intensive and difficult to scale. We investigate whether instruction-tuned large language models (LLMs) can support multidimensional analysis of collaborative sensemaking without task-specific training, and whether structured knowledge-state information improves model inference. We evaluate two mid-size LLMs on 23 richly annotated, expert-labeled episodes across prompting conditions that vary definitional scaffolding, reasoning mode, and turn structure. Without reasoning, models tend to overpredict successful sensemaking; reasoning-enabled prompting improves identification of unsuccessful cases. Knowledge-state diagnostics provide additional grounding, improving detection of unsuccessful sensemaking and increasing agreement with expert annotations. No single configuration performs best across all sensemaking dimensions, underscoring the multidimensional nature of the task.

Post-training quantization (PTQ) is a practical approach to reducing the memory and computational footprint of large language models (LLMs) without retraining. GPTQ-based methods have become the de facto standard, yet they suffer from two complementary limitations. Methods with local, layer-wise objectives lack global supervision; while methods with global objectives fix their Hessian estimates at the start and ignore first-order gradients, so their guidance grows stale as quantization proceeds. This paper presents G$^2$PTQ, a unified PTQ framework with Generalized Gradient Compensation that integrates both first- and second-order information under a globally supervised, block-wise optimization objective. By refreshing gradient and Hessian estimates before quantizing each Transformer block, G$^2$PTQ avoids the staleness of prior global methods. Furthermore, to stabilize the exact first-order compensation, we introduce a trust-region scaling mechanism that dynamically bounds the gradient step to prevent exploding weight updates. Finally, we derive efficient implementations for block-wise Hessian approximation and exact gradient compensation. Experimental results on various model families and bit-widths demonstrate that G$^2$PTQ enables better alignment with the full-precision model, outperforming state-of-the-art baselines. Code is available at: https://github.com/G2PTQ/G2PTQ.

Transformers have become a central architecture for in-context learning (ICL), particularly through their state-of-the-art performance in large language models. This success motivates understanding how transformers exploit task-relevant structure in geometrically heterogeneous data. However, existing nonparametric ICL theory has largely focused on Euclidean domains or single-manifold models. To address this gap, we study the prediction problem under unknown local geometry, modeled by sample size-dependent mixtures of manifolds with heterogeneous dimensions, smoothness, and sampling masses. Under local separation and small-perturbation conditions, we establish a minimax lower bound capturing the aggregate difficulty of the components and construct an oracle tangent local-polynomial estimator with a matching upper bound. This estimator is connected to a structure-informed, two-stage softmax transformer with a geometric preconditioner and chartwise reduced local-polynomial solvers. The transformer achieves negligible approximation error relative to the minimax rate with logarithmic depth and polynomial size. Finally, we derive an in-context generalization bound for near empirical risk minimizers over this class. Together, these results identify conditions under which the resulting predictor exploits local geometry and attains the aggregate minimax rate.

Large language models (LLMs) are increasingly involved in family education, yet HCI has not systematically explained the educational interactions that emerge around them. This scoping review analyzes 53 HCI studies from 6,540 records across 19 venues. Using activity theory and AODM, it relates participants and educational objects to mediation, labour, and rules. We find that the literature centers on child--parent interaction and on language, AI literacy, and relational learning. The introduction of LLMs enabled conversational, embodied, and spatial systems to generate support from the context of an unfolding interaction. LLMs redistributed educational labour, while family and institutional rules left parents and professionals responsible for interpreting outputs and deciding how they entered practice. Evidence across families and educational purposes remains limited, especially on sustained personalization, repair labour, and how families negotiate authority and rules. The review offers a framework explaining how LLM capabilities become organized through family participation.

AI-driven scientific discovery accelerates research by autonomously developing solutions and designs. Large language model (LLM) agents support this process through iterative generation and evaluation. Yet these iterations alone do not ensure cumulative progress or establish which directions to pursue next. Costly evaluation further constrains the scope of exploration. Neural architecture discovery brings these challenges together, coupling open-ended design with resource-intensive experimentation. We introduce EvoTreeNAD, a genealogy-guided evolutionary algorithm that constructs trainable architectures without a supplied seed or a hand-specified search space. Starting from an empty root, it grows a persistent genealogy in which each new node represents a complete architecture. Top-percentile values computed from each node and its descendants guide lineage selection. Using the selected design history, an Idea Agent proposes a variant and a Code Agent implements it. Each evaluated variant becomes a child node, expanding the genealogy while providing evidence for subsequent lineage selection. Our theoretical analysis establishes the existence of stationary variation regimes as the genealogy grows. Under specified variation assumptions, sustained top-percentile family values quantify the probability of generating high-reward architectures in these regimes. EvoTreeNAD discovers architectures that outperform the compared NAS and NAD baselines, achieving CIFAR-10/100 test errors of $2.05{\pm}0.06%$ and $15.09{\pm}0.22%$. On all six MedMNIST-v2 tasks, the discovered architectures surpass the strongest listed baselines. A controlled CIFAR-10 study further shows that EvoTreeNAD outperforms direct generation, best-of-$N$ greedy continuation, and full-family-mean routing.

Chain-of-thought (CoT) instructions can distort multiple-choice VLM evaluation when a scorer appends a reasoning cue but reads answer-label logits before the model generates any rationale. We call this CoT-prefix scoring. On ScienceQA, Qwen2.5-VL-7B drops from 80.76% to 45.48%, and across five option-content permutations 93.54% of CoT-prefix predictions select the first slot. Condition-matched linear probes recover 78.94% from the same hidden states, while free generation restores 75.24%, showing that the answer often survives the prefix and the immediate readout fails. Vocabulary and layer diagnostics explain the mismatch: probability mass moves toward continuation tokens, while answer information remains linearly accessible in late layers. The effect recurs with varying severity across datasets and models, though not universally. These results show that CoT-prefix scoring can confound model knowledge with an evaluation-interface mismatch and should be avoided unless the requested and scored output events are aligned.

Interpretable machine learning for Large Language Models (LLMs) increasingly relies on sparse probing methods that identify small sets of neurons claimed to detect and causally influence behaviors such as factuality recall, safety alignment, and hallucination. These claims have important implications for model auditing and behavioral steering, yet they are rarely tested against known failure modes of $L_1$-regularized probing in correlated, high-dimensional feature spaces. We propose a five-step diagnostic protocol covering feature correlation, bootstrap stability, sparse versus dense ranking disagreement, intervention baselines, and cross-dataset evaluation as a minimum standard for sparse-neuron localization claims. We investigate prior work using our proposed approach, specifically on H-neurons using open-source LLMs across TriviaQA, BioASQ, and NQ-Open datasets. Our results demonstrate detection replicates across both models and datasets, and exceeds the original reported AUROC gaps for TriviaQA and BioASQ datasets. Gemma 3 4B consistently outperforms MedGemma 4B on matched datasets, with AUROC gaps of +0.311 versus +0.235 on TriviaQA, +0.474 versus +0.455 on BioASQ, and +0.128 versus +0.112 on NQ-Open respectively. Causal validation at $n = 500$ with five random seeds shows statistically significant effects beyond random same-layer baselines. At the same time, the diagnostic results indicate that the selected neurons are not uniquely localized. Across the three Gemma 3 4B settings, 19 of 22 selected H-Neurons have Pearson $|r| > 0.7$ with other features, bootstrap selections show only moderate stability, and sparse and dense rankings overlap only weakly. Our findings show that sparse predictive structure can coexist with non-unique neuron selection. Routine diagnostic validation is necessary to distinguish detection claims from localization claims in mechanistic interpretability.

LiDAR provides precise geometric information for spatial perception tasks such as object detection in autonomous driving and outdoor robotics. However, recognizing and localizing individual objects is not sufficient to answer questions that require composing spatial relations and grounding the intended target. Motivated by recent advances in large language models (LLMs) for autonomous driving, we leverage their language priors to interpret complex spatial questions and ground the referred target in LiDAR geometry. To support this spatial grounding capability, we introduce SpatialLiDAR-QA, which combines single- and multi-step relational grounding with complementary spatial understanding tasks. We further propose SpatialLiDAR-LM, which aligns LiDAR point features with an LLM and grounds target coordinates through language-conditioned, position-aware proposal retrieval and local point refinement. This design derives target coordinates directly from local LiDAR geometry rather than through textual language decoding. Experiments demonstrate substantial improvements over representative LiDAR--language models and multi-camera VLMs on precise coordinate prediction tasks. Our dataset and model training code will be publicly released.

Multimodal large language models (MLLMs) achieve strong performance on visual reasoning tasks, yet remain prone to hallucinations and over-reliance on language priors, often generating answers without adequately using task-relevant visual evidence. Existing approaches primarily improve reasoning through reasoning-oriented supervision or inference-time strategies. In this work, we study a complementary question: can multimodal reasoning be improved by strengthening implicit visual grounding without directly supervising the reasoning process? Motivated by the functional specialization of attention heads, we investigate whether reasoning can be improved by guiding only the heads most responsive to visual evidence grounding. We propose Selective Probability Mass Concentration (sPMC), a training framework that identifies grounding-responsive heads and selectively regularizes their text-to-image attention. sPMC treats normalized attention over visual tokens as a spatial probability distribution and encourages the probability mass to be assigned to semantically relevant regions using segmentation-derived spatial priors. Adaptive Head Selection restricts this guidance to visually responsive heads while leaving the remaining heads unconstrained to preserve their complementary functions. Across 6 multimodal benchmark suites, sPMC achieves an average zero-shot improvement of 3% and gains of up to 11.3% across multiple MLLMs while regularizing only 3%-15% of their attention heads. These results demonstrate that targeted guidance of sparse and implicit visual evidence pathways can directly improve multimodal reasoning.

Advances in language modeling have been driven by scaling pretraining on ever more data. Yet, the training data is still largely curated on the model's behalf. A more general approach to pretraining would let the model learn to generate the data most useful for its own improvement. This would provide an effectively unbounded source of training data, limited by compute rather than human knowledge. We introduce Self-Play Pretraining with Zero Data, an initial proof-of-concept towards realizing this vision. Our procedure casts synthetic data generation as a search over the space of all computable structure, taking inspiration from Solomonoff induction. Starting from random initialization, two models learn in tandem: a generator proposes programs interpreted by a universal Turing machine, generating byte sequences, while a learner autoregressively predicts these byte sequences. The learner is trained with standard cross-entropy, while the generator is trained with reinforcement learning to produce sequences at the frontier of the learner's capabilities, yielding an adaptive curriculum. A universal Turing machine gives us a search space over all computable data-generating processes, imposing little domain-specific structure, and self-play searches over this space for useful training data. We test whether zero-shot performance on natural data improves predictably with self-play compute; this is a clean test of transfer since neither generator nor learner is trained on natural data. Across several natural datasets, zero-shot loss exhibits predictable scaling in compute. The models also exhibit in-context learning, and discover recognizable mathematical sequences during training.

The English Resource Grammar (ERG) is a hand-written computational grammar of English. Given a sentence, its processor, ACE, produces a formal meaning representation called Minimal Recursion Semantics (MRS): a graph of the sentence's predicates and their arguments. The grammar is bidirectional and can also turn an MRS back into an English sentence. \citet{hajdik2019} used the ERG's treebank to build a benchmark for that generation task, MRS to text, and trained sequence-to-sequence models to solve it. The parsing task, text to MRS, can be tested on the same sentences. We reconstruct their 10K-sentence test split, and score two large language models, Claude Sonnet~4.5 and Claude Opus~5, in both directions against their trained systems and against ACE, with no task-specific training. Given an MRS and three examples, Opus writes the sentence at 76.3 BLEU, ten points above their system trained on 72k pairs (66.1 BLEU), and comparable to their system trained on a million extra pairs (77.2 BLEU). Sonnet scores 65.7 BLEU, and letting it choose among ACE's own candidate sentences lifts it to 69.6, while a pooled judge that keeps Opus's own sentence among the candidates adds 0.6 points (77.0 BLEU). In the parsing direction, however, the models fall far behind ACE: asked for the MRS of the same sentences, they reach 57.2 (Sonnet) and 65.5 (Opus) F$_1$ on the graph's predicates and arguments against 91.0 for ACE, and exact-match the gold on about 1% of sentences. We characterize the failure modes for the parsing tasks, and conclude that a generation score alone does not show that models understand formal semantic representations.

Large language models increasingly operate as persistent assistants in user-facing, shared-session, and tool-augmented settings. When users disclose sensitive information during an active conversation, that information may remain behaviorally recoverable through later prompts even after the dialogue shifts to unrelated topics. We introduce PrivDrift, a benchmark for auditing whether user-disclosed secrets remain recoverable after conversational topic drift and persuasion-based probing. PrivDrift contains 1,000 controlled multi-turn dialogues with seeded secrets, content-dense drift turns, and standardized extraction probes. Across three LLMs with extended context windows, dialogue-level hybrid leakage remains substantial, ranging from 38.7% to 54.6%, and varies strongly by model, secret type, and persuasion intensity. Within the tested drift window, additional topic drift does not reliably reduce leakage, suggesting that privacy risk in active LLM contexts should be evaluated as a persistent behavioral failure mode rather than only as training-data memorization or immediate jailbreak behavior.

Recent advances in paraphrase detection reveal a fundamental trade-off: large language models achieve high accuracy but require high computation, while efficient Siamese-BERT variants offer practical scalability with reduced transparency in rationale generation. We present R-DEIM Net, a 76M-parameter dual-expert architecture exploring whether moderate-scale models can achieve competitive accuracy on paraphrase detection while enabling human-readable rationale generation. The architecture combines two specialized components: an Interaction Expert that captures token-level similarity patterns through multi-scale 2D convolutions and attention head allowing variable input length, and a Reasoning Expert that uses a Flan-T5-small decoder to generate rationales as auxiliary supervision. Rather than re-encoding generated text, we extract and pool decoder hidden states as complementary features for classification. On the Quora Question Pairs dataset, R-DEIM Net achieves 90.07% accuracy and 90.16% F1-score via 10-fold cross-validation. This represents competitive performance with strong transformer-based baselines (e.g., MFAE BERT: 90.54% accuracy) and recent large language model based approaches (LLaMA-70B) while using a substantially smaller parameter budget. The model generates rationales alongside predictions, providing potential for auxiliary human-readable descriptions.