language model engineers

Abstract: Continuous diffusion language models (CDLMs) offer an alternative approach to parallel language generation; however, most existing studies directly use transformer architectures from either causal large language models (LLMs) or Diffusion Transformers (DiTs) for image generation, which were not optimized for CDLMs. In this paper, we conduct a systematic study on Transformer architectures for CDLMs and introduce Language Diffusion Transformer (LangDiT), combining feature and query-key normalization, wider noise conditioning, likelihood-shortcut gradient routing and calibration, and confidence-gated training self-conditioning. LangDiT significantly reduces generation perplexity in matched LangFlow comparisons, establishing a new state of the art in continuous language generation across different model formulations. Using LangFlow as a representative CDLM, LangDiT as a plug-and-play backbone leads to a new state-of-the-art model with a generation perplexity (GenPPL) of 21.72 using 1024 NFEs and 44.72 GenPPL using 128 NFEs on the unconditional generation of 1024-token sequences on the OpenWebText dataset. When adapted to Plaid-style CDLM formulations, LangDiT results in a new leading GenPPL of 11.33 using 1024 denoising NFEs, and 27.51 GenPPL using 128 denoising NFEs. These gains are achieved while retaining each model's underlying representation, training objective, and inference sampler, isolating transformer architecture as a fundamental source of improvement. Our study identifies architectural design as an important source of improvement in continuous diffusion language models.

Abstract: Looped Transformers improve parameter efficiency by repeatedly applying the same Transformer layers. However, generating each token still involves sequential computation and repeated access to model weights, leading to high decoding latency. Recent pipelined self-speculative decoding methods reduce this delay by drafting tokens from early-loop predictions and asynchronously advancing tokens at different recurrent depths in the same batch, but they do not fully exploit early-loop features for longer lookahead and still spend computation verifying incorrect speculative branches. We introduce \method, a speculative decoding framework for looped transformers that enables earlier drafting of future tokens and assigns complementary roles to different recurrent depths. A lightweight lookahead drafter (LAD) predicts multiple future tokens with early-loop states, and deeper-loop feedback helps decide which draft branches to continue and which to pause, while the full-depth model makes the final token decisions. LoopAhead achieves up to 7.82x decoding speedup over autoregressive decoding while preserving the target model's output distribution.

Abstract: Addressed conditional memory provides a local surface for post-hoc knowledge injection, but a successful write need not transfer to a new question or preserve untargeted behavior. This study combines a controlled GPT-2 comparison with production-memory experiments on Qwen3.8-Flash-Next-FP8 and DeepSeek-V4.1-Flash. Newly initialized modules attached to frozen GPT-2 are tested in a 2 x 2 design: row-only versus row-plus-projection updates, each with or without fixed module-residual anchoring. All arms reproduce 50 trained completions in each of three sampling seeds. Natural-question target-string hits average 5.3% for both row-only arms, 28.7% for unanchored projection updates, and 22.0% with anchoring, while strict answer scores remain zero. Anchoring lowers measured drift; lexical hits do not establish entity-specific retrieval. In the archived Qwen experiments, gradient writes make the designated first answer token top-ranked for 132 independently reset targets, comprising 93 new promotions and 39 already top-ranked targets. A separate matched 24-target comparison gives 13 new promotions with Adam versus two with SPSA at equal iteration counts but unequal computational procedures. An 80-row case includes saved continuations and a replay record, while reformulations bound its scope. Six DeepSeek zero-order cases yield one new target-token promotion and one regression. These results support bounded writing feasibility and separate trained expression, question-form behavior, and preservation; first-token rankings do not establish complete-answer success or universal conditions for writable memory.

Abstract: On-policy distillation (OPD) improves a student language model by training on its own generated trajectories with supervision from a stronger teacher. However, when conditioned on student-generated prefixes, the teacher can favor invalid continuations that violate the question's premises, causing OPD to encourage erroneous reasoning steps. As a remedy, we introduce Condition-Contrastive Extrapolated On-Policy Distillation (CCE-OPD), which makes teacher supervision sensitive to question conditions. CCE-OPD incorporates (1) a fixed-prefix comparison in which the same frozen teacher evaluates the complete question and a counterpart with a selected condition masked, and (2) a bounded log-probability contrast that reweights the teacher distribution into an auxiliary distillation target. On 200 diagnostic reasoning states, condition contrast reduces the rate of preferring invalid continuations from 27.0% to 16.0%. Extensive experiments across six mathematical benchmarks show that CCE-OPD improves avg@8 over OPD by an average of 1.60 and 1.95 percentage points when distilling Qwen3-8B into 0.6B and 1.7B students, respectively. It also achieves the strongest results on four transfer benchmarks.

Abstract: Sparse KV compression reduces long-context inference costs by retaining selected historical states, yet these states typically keep their coordinates from the full history. Under Rotary Position Embedding (RoPE), gaps left by evicted tokens continue to affect query-key matching, potentially limiting the use of retained evidence. We study post-selection positional geometry on a fixed selector-produced cache, examining how positional organization affects the use of retained information and which coordinate and continuation rules improve answer quality. To reorganize this geometry, we propose RoPERelay, which assigns compact, order-preserving RoPE coordinates, relocates cached key phases through inverse rotation and re-rotation, and synchronizes the current query and subsequent generation. It reuses existing K/V without renewed selection, historical re-prefill, or additional cache budget. Across 36 comparisons spanning two models, six sparse selectors, and three long-term memory tasks, RoPERelay improves answer quality in 32 cases. For Llama-3.1-8B-Instruct with BM25 on LongMemEval-S, accuracy increases from 23.6% to 35.4%. Measured complete-request time overhead is approximately 1%-3% in the evaluated configurations. These findings establish positional organization as a distinct optimization dimension for sparse KV caches after content selection.

Abstract: Parallel drafting reduces the drafting overhead of speculative decoding for large language models (LLMs), but its gains remain limited by the accepted prefix length. Even when the correct token is present in the candidate pool, a single early selection error prevents subsequent predictions from being used. We propose DRelay, which uses global information from the entire draft block to perform prefix-aware selective repair of candidate selections before target-model verification. DRelay bases its decisions on candidate correlations and the selected path: a global reader extracts predictive information across positions for each candidate. While a causal selector combines candidate-level information extracted by the global read with the tokens selected at preceding positions to determine whether the native choice at the current position is consistent with the global evidence and the selected prefix. It then decides whether to retain or replace the token, thereby repairing early errors and extending the accepted prefix. We further jointly train the draft backbone and the selector, combining candidate-support learning with a repair objective, while weighting the repair loss according to each block position's potential contribution to the consecutive accepted prefix. Across eight diverse benchmarks on an H800 GPU, DRelay consistently improves both average acceptance length and end-to-end decoding performance over DFlash, Domino, and DSpark. Under SGLang serving, DRelay improves average end-to-end speedup over DFlash, Domino, and DSpark by 14.7%-16.8%, 8.7%-9.3%, and 8.1%-9.3%, respectively.

Abstract: Mamba-style and hybrid language models compress their past into a fixed-size recurrent state that is rewritten at every generated token. Storing this state in low precision saves memory bandwidth, but every rounding error is fed back into the next update and can accumulate over long generations. Production systems round the state stochastically; we ask which rounding rule such caches should use. We find that a deterministic golden-ratio Weyl dither, which needs no random numbers, consistently brings the quantized model closer to the full-precision one than stochastic rounding, across pure and hybrid models, storage formats, and long decoding horizons, at no extra cost. Round-to-nearest behaves differently: because it discards small updates, its error keeps growing, so it can look best in short evaluations yet falls far behind over long generations. A discrepancy analysis explains this ordering, and we document implementation pitfalls that silently remove the benefit.

Abstract: On-policy self-distillation (OPSD) has attracted growing interest as a promising approach to improve the reasoning ability of language models. Without external rewards nor a separate stronger teacher, the self-teacher with privileged information could provide dense signals on student's trajectories. However, its behavior in language reasoning remains unclear, with reported outcomes ranging from modest gains to behavioral collapse. In this work, we diagnose OPSD for mathematical reasoning across models spanning 0.6B--8B parameters. We conduct controlled experiments and token-level analyses to fully delve into OPSD. We point out that teacher's signal is shaped by reasoning-mode alignment and the complete teacher prefix, rather than by privileged semantics alone. OPSD improves reasoning only in narrow compatibility regimes. Otherwise, it produces ineffective length growth, stable degradation, or behavioral collapse. Token-level analysis shows that teacher's signal is not stable and does not predict downstream performance. Based on these results, we argue that OPSD is a sensitive algorithm rather than a generally reliable reasoning-improvement post-training method.

Abstract: Test-time reasoning systems often respond to failure by restarting or revising the latest step, even when an earlier decision caused the error. We introduce CORE, a search controller that requests a certified conflict core from a verifier, backjumps to the latest decision in that core, and caches the conflict to avoid repeating it. Under sound verification, finite branching and depth, and exhaustive proposals, the uncapped search is complete and never prunes a valid solution. On 2,000 planted graph-coloring instances with matched proposals and an exact verifier, CORE reduces median verifier calls by 39.8% at 30 variables and 35.0% at 36 variables relative to chronological repair; caching further improves on backjumping alone. Across five reasoning tasks, CORE achieves 75.9% mean success with Qwen2.5-7B-Instruct and 84.2% with Qwen3-8B, compared with 72.5% and 81.8% for Tree of Thoughts. It also uses fewer verifier calls and generated tokens on both backbones. These results show the value of using certified failure explanations to direct language-model search.

Abstract: Discrete diffusion models and flow matching have emerged as powerful frameworks for generative modeling over discrete state spaces, yet efficient few-step generation remains a fundamental challenge. In this work, we introduce the Discrete Average Generator, a principled extension of MeanFlow to Continuous-Time Markov Chains (CTMCs). Analogously to how MeanFlow defines an average velocity field over a time interval in continuous spaces, we define an average generator as the normalized increment of the transition kernel over a time interval. We show that this average generator satisfies a self-consistency identity, which provides the foundation for our training objective. We further develop training strategies that align with the standard training paradigm of diffusion language models while keeping the resulting objective tractable. When projected onto per-coordinate marginals, the self-consistency identity admits a closed-form expression, enabling efficient training and inference. In Potts model simulations, our objective reduces the total variation distance of the $K$-step sampler by up to 67%. On OpenWebText, our method achieves the lowest generative perplexity among the evaluated methods for 8 to 64 sampling steps while enabling a $16\times$ acceleration, and achieves comparable performance to existing methods on ImageNet.

Abstract: Large language models (LLMs) excel at token-level generation but may learn undesirable abstract semantics and lack comprehensive perception. LLM-JEPA mitigates this by aligning different views of the same underlying knowledge via a joint-embedding predictive architecture (JEPA). However, strong alignment does not necessarily lead to accurate, stable predictions. To address this, we propose ER-JEPA, which adds an episodic replay path to LLM-JEPA. ER-JEPA stores training pairs in a memory. At each step, it stores and retrieves relevant data to provide additional supervision. This enables learning from both the current batch and stored training pairs, providing additional supervision for token prediction and representation alignment. Experiments across multiple datasets (NL-RX, GSM8K, Spider, and NQ-Open) demonstrate that ER-JEPA consistently outperforms LLM-JEPA.

Abstract: On-policy distillation (OPD) reduces train-test state mismatch by training a student on its own generated trajectories, but weak students may visit teacher-misaligned prefixes where supervision is less representative. We introduce SAKI (Supervision Allocation with KL-constrained Interpolation), which combines a KL-constrained teacher-guided rollout with maximal coupling and reuses realized accept/correction events to route token-level supervision. Accepted positions retain sampled-token reverse-KL supervision, while correction positions receive direct supervision on the teacher's highest-probability token. Under maximal coupling, the correction probability is exactly TV(p_t, q_t), so the same trust-region radius controls rollout deviation and upper-bounds intervention and specialized-supervision frequency. We further implement an engine-resident speculative verifier that preserves the exact-q trajectory distribution and coupling semantics while improving matched-workload rollout throughput by 4.22x. Across seven mathematical reasoning benchmarks, SAKI improves the matched teacher-guided baseline in Mean@8 and Pass@8 for both 1.7B and 0.6B students. Placement controls and fixed-prefix analysis further support correction-triggered routing as a conflict-adaptive supervision signal.

Abstract: State Space Models (SSMs) compress sequence history into a bounded recurrent state, making the resulting memory law a central architectural choice for long-context performance. Most modern SSMs rely on ODE-based dynamics that lead to exponential forgetting, limiting their ability to retain information over broad temporal ranges. We introduce FRAC, a selective SSM architecture derived from fractional dynamics that replaces this exponential decay with power-law long memory. To make fractional dynamics practical, FRAC approximates the heavy-tailed target kernel with a finite-state, log-spaced sum of exponential modes. This construction turns fractional memory into an efficient recurrent module with parallel training and prefill, while retaining bounded-state autoregressive decoding. Extensive experiments, including 1.3B-parameter language modeling, demonstrate that FRAC consistently improves long-context performance over state-of-the-art SSM baselines while staying competitive on short-context. These results show that fractional dynamics provide a practical and effective prior for long-context SSMs.

Abstract: Large language models can learn abstract rules from just a few in-context examples, but how their internal mechanisms activate as examples accumulate is not well understood. We trace a three-stage symbolic reasoning circuit (abstraction, induction, retrieval) across shot counts in three model families and find it is detectable and functional well before the model achieves high accuracy. Per-head causal contribution grows up to 8x from 1- to 10-shot, and cross-shot activation patching raises accuracy from 1% to 56% at 0-shot and 17% to 88% at 1-shot. Function vectors scaled and injected at 0-shot rescue accuracy up to 86%, largely substituting for the induction stage but depending critically on an intact downstream retrieval stage. The infrastructure for abstract rule-following is present in the weights before any demonstrations; in-context examples, function vectors, and related interventions appear to supply input to the same latent circuit.

Abstract: Parallel speculative drafting generates multiple candidates in one backbone pass, but independent token selection can produce inconsistent continuations that shorten the accepted prefix. Existing methods mostly leave conditional decoding to a lightweight module after the backbone, which limits the flow of predecessor information to successors. Our analysis of DFlash shows that early positions already form recoverable predictions in shallow layers, and that accurate adjacent predecessors help successors more when they enter earlier. We therefore propose DSpine, a drafter with causal conditioning injection throughout the backbone: at every layer, gated adjacent injection writes each predecessor's predicted feature into its successor, so the causal conditioning chain unfolds over network depth while all positions update in parallel. A unified transfer space built from the target model's output embeddings unifies layer-wise injection with predecessor-conditioned decoding, and layer-wise output-embedding supervision promotes the formation of predicted features in shallow layers. Fused kernels and a transition cache execute both efficiently in parallel within SGLang. Across seven math, code, and chat benchmarks, DSpine achieves the longest acceptance length at both temperatures on Qwen3-4B and Qwen3-8B. At temperature zero on Qwen3-8B, it raises the seven-benchmark mean from DFlash's 3.77 to 4.82 (+27.8%); in SGLang serving tests, it delivers 23.3% higher throughput than DFlash on average.

Abstract: Large language models (LLMs) reliably perform entity copying, in which a model copies tokens referring to an entity, termed entity tokens, from the prompt into its output to answer a question. Although entity copying is straightforward for most LLMs, existing research does not provide a systematic account of which layers specialize in this fundamental task or how other tokens in the same sequence, termed context tokens, influence the model's ability to copy the entity tokens. To address these questions, we conduct experiments on Qwen3-8B using two novel methods: genie-in-a-bottle, which controls exactly which layers can participate in an entity-copying task, and attention lobotomy, which cuts off specific tokens' attention to entity tokens without affecting the remaining attention distribution. We find that two distinct groups of layers in the second half of the model are both necessary and sufficient for entity copying. Moreover, in addition to the decoding position's attention to entity tokens, context tokens' attention to entity tokens also proves necessary for copying the exact tokens, even though context tokens do not store entity information themselves unless they satisfy particular semantic properties. Our findings establish the critical role of late layers in entity copying under the guidance of context tokens, calling for future work on how models propagate and consume entity information.

Abstract: On-policy distillation (OPD) transfers knowledge between language models through teacher supervision on student-generated trajectories. With different tokenizers, a single teacher token may require multiple student tokens to generate, creating intermediate states where the event is entered but not yet completed. Existing cross-tokenizer methods align tokens or text spans to construct comparable prediction targets. We study a complementary problem after partial generation: once the student produces a prefix of a teacher token, multiple next tokens may complete the same remaining bytes, but the teacher only specifies the required completion rather than how probability should be divided among these valid continuations. We introduce Event-Set Completion Distillation (ESCD), which complements cross-tokenizer probability alignment with completion-set supervision. ESCD aggregates prefix-related teacher events and supervises the total probability of byte-compatible one-step student completions, avoiding tokenizer-dependent probability splits among individual tokens. The method reuses student trajectories and predictions, requiring neither additional rollouts nor changes to the student vocabulary. Experiments demonstrate consistent gains in mathematics, code, and scientific reasoning across model families and tokenizers, extending to large-scale MoE distillation from a 1T teacher to a 35B student. Local analyses show that retaining completion sets better matches the reference supervision, while one-step completion covers over 99% of observed compatible teacher mass after partial event entry in the studied tokenizer pairs. These findings support event entry and event completion as complementary supervision targets for cross-tokenizer knowledge transfer. Code will be released on GitHub.

Abstract: Large language models are trained on heterogeneous data mixtures, where different knowledge domains require both shared knowledge and specialization. Existing modular approaches typically impose explicit components or discover modules through interpretability analysis after training. In this work, we propose CORTEX, a learning dynamics-inspired framework that learns internal modularization within dense language models. CORTEX partitions trainable matrices into parameter groups and learns module assignments from domain-conditioned gradient and cross-domain gradient similarity. We introduce the selective lesion score and module-domain mutual information to characterize the target-domain lesion effects and alignment, and analyze how module assignment affects the trade-off between assignment bias and update magnitude. Experiments with 160M, Qwen3-8B, and Qwen3-32B backbone models show that CORTEX achieves the highest synthetic-domain exact match and largest average perplexity reduction, while remaining competitive on real-domain evaluations and forming identifiable modules.

Abstract: Pretrained autoregressive Transformers represent a large sunk investment in compute. Existing conversion methods reuse that investment by changing either the architecture (attention to recurrence) or the objective (next-token prediction to denoising), never both. We convert Qwen3 teachers at 1.7B and 8B into attention-free, bidirectional, gated-delta-rule diffusion students in three stages, so that each capability can be traced to the stage that kept or lost it. Language modeling transfers only partially and in-distribution; in-context retrieval does not transfer. On a multi-query recall probe where the teachers score 0.34-0.58, both converted students score 0.000, and diffusion pretraining alone does not restore retrieval. A retrieval curriculum in the final stage, which gradually lengthens the gap between a key-value table and the queries that address it, restores it only stochastically: on a fixed schedule, one seed in three learns to retrieve. Advancing the gap only while a running accuracy estimate stays above a threshold works for all three of those seeds, holds on real text, and carries unchanged to 8B, where two of three seeds succeed. The third had not learned within its fixed 16k-step budget: retrieval switches on abruptly at a seed-dependent step (6.5k and 11k in the other two), so a fixed budget can cut a late run off. One boundary survives every intervention: every model that learns retrieval scores 0.000 on tokens that never appeared in a retrieval episode, and an arm that resamples the key and value tokens every batch shows this is a coverage limit, not memorization of particular bindings. Separately, we convert a 7B code model into a 3:1 recurrent-attention block-diffusion hybrid over 85k steps and report two negative training results.

Abstract: Looped language models separate computational depth from parameter count by repeatedly applying the same transformer block. Adapting these models requires a shared update that remains effective as hidden states evolve throughout the recurrent computation. Our empirical analysis reveals a pronounced late-loop bias in standard low-rank adaptation (LoRA): the shared update is more effective at later loop positions. This imbalance motivates training shared updates under varying combinations of their applications. Randomly omitting adapter applications alone, however, does not improve task performance; it reduces expected update strength during training while leaving inference unchanged. We introduce Loop Dropout, which couples stochastic masking of adapter applications with inverse-survival rescaling to preserve expected update strength and promote effective adaptation across loops. Extensive experiments demonstrate improved mathematical reasoning across model sizes, adapter ranks and training recipes, with benefits extending to general instruction tuning and code generation. Loop Dropout outperforms existing LoRA variants and adapter regularizers, while further analysis shows stronger early-loop adaptation. Every backbone loop remains active, and inference applies the adapter at all loops using standard LoRA without additional trainable parameters or inference computation.

Abstract: Residual completion augments query-aware sparse attention by estimating the contribution of tokens omitted from the exact sparse computation. We ask whether improving a layer's attention-output reconstruction on the same incoming Q/K/V and selected support necessarily improves the fidelity of the final model output. We study training-free RESA and learned Top-K+$φ$ with frozen backbone language models. A prespecified single-layer screen yields two Qwen3-0.6B/Multi-LexSum interventions for which direct-runtime measurements show positive prespecified request-aggregate local reconstruction gain but worse final KL fidelity than the corresponding all-abstain Exact Top-K baseline on both discovery and prompt-token-disjoint holdout requests. Exact restoration at the same layer instead improves final fidelity, showing that the reversal is specific to approximate completion in these cases. In complementary multi-layer experiments, a task-independent local diagnostic often repairs the tested completion estimators, although the repaired models do not consistently outperform Exact Top-K. Together, these results show that better local reconstruction need not translate into better final-model fidelity.

Abstract: Most continuous diffusion language models process one latent position per token at each sampling step, making generation expensive. Two-stage methods lower the cost by reducing the latent length, but they fix the compressed embedding space before training the diffusion model. Embeddings from the fixed space can be difficult to model with diffusion and decode reliably into tokens, which limits generation quality after compression. To address this problem, we introduce JPEG-DLM (Joint-embedding Prediction for Efficient Generation with Diffusion Language Model), which jointly trains a compressor, a flow matching model and a decoding module. With joint-embedding prediction, JPEG-DLM learns compressed embeddings that are more structured, easier to model with diffusion and reliably decodable into tokens. JPEG-DLM achieves the lowest mean Gen-PPL and highest throughput among recent diffusion and flow models on LM1B and OWT. At a compression rate of 0.5 on OWT, it reaches a Gen-PPL of 34.52 and approximately 2.3 times ELF's throughput. These results suggest that jointly learning compressed embeddings offers a promising path toward efficient diffusion language modeling. Code will be released soon.

Abstract: Context-aware dynamic trees allocate the speculative decoding budget according to draft path probabilities, adapting their depth and branching to the current context. Under stochastic decoding, however, we find that this structural advantage does not always compensate for the acceptance gains of random sampling paired with advanced verification, and such dynamic trees can fall behind sampled chains in some settings. These trees grow their topology from the candidates themselves, so the tokens submitted for verification are typically the deterministic high-score tokens selected during construction. This coupling is not inherent: once the topology is fixed, its nodes can be repopulated by sampling, allowing dynamic trees to retain their structural advantage while also benefiting from random sampling and advanced verification. Diffusion-based drafters make this practical, as their parallel outputs or lightweight conditional corrections allow candidates to be regenerated cheaply after the complete topology is known. We introduce Tsubame, a two-pass tree speculative decoding framework for diffusion-based drafters. The first pass plans and freezes a context-aware topology using draft path scores; the second replays the fixed topology, sampling the tokens that populate its nodes to form the candidate tree for verification. We prove that Tsubame is lossless under compatible sampling and verification strategies. Experiments across three diffusion-based drafters, six datasets, and multiple candidate budgets show that Tsubame improves acceptance length and throughput over deterministic trees, including settings where it reverses their disadvantage against sampled chains.

Abstract: Autoregressive pretraining increasingly draws on heterogeneous data, making it important to understand how a model learns from an individual token. The next-token prediction objective naturally identifies a token's contribution with its own loss. However, each token is not only a prediction target but also context for what follows. Using controlled corruption, we decouple these two roles and find a reversal: making a noisy token easier to predict reduces its damage as a target but increases it as context. The same decoupling helps explain text generated by language models: generation selects each token by its fit to the prefix, while its role as context is never tested against an independently determined continuation, because that continuation is generated to fit it. At known corrupted positions, acting through the context can reduce damage that removing the token's own loss does not. Understanding and controlling what a model learns from a token therefore requires decoupling its roles.

Abstract: Masked diffusion language models (MDMs) admit flexible generation orders, making the unmasking strategy an inference decision. Existing methods vary in how they prioritize positions, control parallelism, restrict selection regions, revise predictions, or plan future denoising, yet it remains unclear when these choices should change during generation. We study this question through strategy reversals, where an alternative action becomes preferable to a fixed choice. We organize MDM inference into five axes--score, cardinality, region, commitment, and planning--and define adaptation opportunity as the one-step utility advantage of the best candidate action over a validation-selected fixed action. This view shows that adaptation value depends on both the frequency and magnitude of such reversals. Across three MDMs and ten tasks, adaptation opportunities are highly heterogeneous, with some regimes exhibiting concentrated and predictable one-step gains. This motivates selective adaptation: lightweight detectors calibrated on validation prompts identify high-opportunity states, capturing, for example, 56.9 percent of the candidate-set oracle opportunity by adapting only the top 10 percent of states on LLaDA-8B constrained JSON filling. Our transition-level results suggest that state adaptation is most useful when applied selectively rather than uniformly.

Abstract: Reinforcement mid-training lets language models learn internal thoughts from unlabeled text, but efficient token-level credit assignment remains challenging. Existing group-relative methods require costly repeated generation. Learned critics offer single-rollout feedback, but accurate return prediction alone does not ensure reliable policy updates. Our analysis shows how training--inference mismatch and PPO clipping prevent a common offset in advantage estimates from cancelling out, introducing additional update drift. We propose \tfour{}, a twin-critic method that calibrates token-level advantages from a single generated trajectory. After warmup and held-out qualification, the critics provide two advantage estimates, combined using action-dependent weights learned through a conditional-moment saddle-point objective. This objective brings the average advantage at each prefix toward zero, while a signal-retention constraint prevents the correction from erasing the learning signal. Sharing information across text positions avoids repeated sampling of each prefix. Theoretically, we characterize optimal mixing under the signal-retention constraint and establish an upper bound on residual mean-induced drift. Experiments show that, compared with the state-of-the-art critic-free method, \tfour{} improves mean benchmark performance by 7.8\% and reduces mean training-step time by up to 63.4\%.

Abstract: Large language models represent context with a key-value (KV) cache. Caches are model-specific: for the same text, models with different architectures or weights produce incompatible representations. This makes it costly to switch models over a shared context: although the context has already been processed by one model, the incoming model must process it again to build its own cache. We introduce KV-Lingo, a method for translating the KV cache of a source model into one that can be read by a target model. KV-Lingo consists of a collection of linear maps, typically one per layer of the target model, that are applied independently on all tokens' key and value representations. We train these maps using distillation, minimising the divergence between the target model's predictions from its native cache and those from the translated cache. We consider several model pairs spanning multiple sizes and architectures, training one translator per pair on a generic text corpus. The resulting translators preserve strong downstream performance in both small-to-large and large-to-small transfers. Since a switch then costs a linear map and a single decoding step instead of a prefill, replacing re-prefill with cache translation reduces the time to first token after a model switch by 9.6x already on a 64-token prompt for Qwen models on an Apple M3 Ultra, and by up to 29x at 32k context length on an H100. These gains make KV-Lingo particularly useful for dynamic model routing: a context can be processed by one model and handed off to another only when needed, without re-prefilling the shared prefix. We finally show that KV-Lingo can be used for seamless model switching, staying close to re-prefill across repeated switches in our multi-turn evaluations.

Abstract: Personalization requires language models to capture user-specific knowledge from a growing user history. Existing context-based approaches incur increasing inference costs as user history accumulates and rely on separate retrieval or summarization stages, while parametric-based approaches often require reconstructing user representations when new user data is added. We introduce ProTTT, a profile-supervised meta-learning framework for learning semantic user memory. The memory construction starts from a shared initialization and is updated for each user through test-time training on user history, allowing it to evolve continuously as the history grows. However, since test-time training alone does not explicitly encourage the memory to capture semantic user knowledge necessary for personalization, we learn this shared initialization using textual user profiles as supervision, so that test-time training on user history captures semantic knowledge more effectively. ProTTT consistently outperforms both full history ICL and all parametric baselines across diverse benchmarks, while substantially reducing inference cost by compressing user history into a lightweight parameterized memory. Our analysis also shows that profile supervision is a reliable objective for learning semantic user knowledge and that the resulting memory can track and retain evolving user preferences, while remaining robust across different history sizes. Overall, we demonstrate the effectiveness of test-time training for personalization and establish ProTTT as a baseline for continuously evolving user memory.

Abstract: Data selection for language model pretraining faces a fundamental tension between quality and diversity. While quality filtering is empirically effective, it often induces diversity collapse: by favoring texts similar to high-quality reference corpora (e.g., educational or QA-style data), it systematically excludes valuable data from underrepresented domains. In contrast, diversified selection preserves domain balance and encourages robust downstream performance, yet existing methods either focus on coverage-oriented objectives that indirectly enhance diversity, or directly optimize for diversity via costly covariance matrix recomputation that limits scalability. To address these issues, we introduce \textbf{Leverage Score Sampling (Lev)}, which iteratively selects samples that maximally expand the determinantal volume of the embedded data via leverage scores, a computationally efficient criterion that eliminates matrix recomputation and enables scalable selection. Empirically, Lev delivers up to $72\times$ speedup and improves dataset diversity, measured by the Vendi score, by $9.2\%$ over the strong diversification baseline \textbf{DiSF}. On CommonCrawl (CC) web data selection, Lev improves accuracy across seven downstream tasks by up to $1.31\%$ over existing baselines. For domains where robust quality criteria are inherently difficult to define (e.g., code), Lev serves as an effective unsupervised curation alternative: on StarCoderData, the selected subset reduces bits-per-byte by $3.08\%$ over DiSF. Notably, we uncover a cross-domain collapse of quality filtering: CC data filtered by DCLM-fastText fail to retain sufficient code-related content, yielding inferior code performance relative to Lev-selected data. These findings advocate for integrating diversity-aware practices into quality filtering for more effective data curation in language model pretraining.

Abstract: Can the global-workspace account of transformer representations extend to state-space language models? We fit Jacobian lenses to the residual streams and recurrent states of Mamba-1, Mamba-2 and Mamba-3, using the original 1000-prompt recipe. Joint residual--state readouts improve recovery of known intermediate concepts over the residual lens on at least five of six task families in every tested Mamba checkpoint. On Mamba-2, state alone exceeds residual and logit lenses on all six families; a normalised joint readout improves on both components on five. Temporal maps and word-list experiments show earlier content remaining state-readable as residual visibility changes. We also propose sign-guarded steering, which improves target top-five success over coordinate exchange on matched verbal-report trials in five models. Recurrent state alone supports this verbal access. These gains do not extend consistently to relational answers: guarded edits often output the edited concept itself, and Mamba-3's joint edits can disrupt successful state-only redirection. Recurrent state thus provides a complementary carrier of workspace content, whose recovery, persistence and causal uses require separate measurements.

Abstract: The quadratic complexity of dense self-attention remains a central bottleneck for long-context language modeling. Many efficient alternatives address this cost by deciding in advance where attention should be sparse or local. We argue that attention approximation should instead be approached as a geometric problem, with the relevant interaction geometry learned from data: natural-language dependencies are input-dependent and difficult to prescribe in advance, so the model should learn where positional relevance can decay and where broader interactions must be preserved. We introduce Mixture of Semantic Attention Regimes (MoSAR), which learns such an adaptive, controlled-decay geometry over query--key interactions. Input-conditioned query and key routers, applied after positional encoding, select mixtures over short, medium, and global regimes, inducing a continuous distance-dependent attention field rather than a fixed sparsity pattern. This geometry is learned during training and can subsequently be discretized through top-1 routing. In controlled pre-training experiments with matched 500M-parameter models, MoSAR learns a substantially lower-reach attention geometry without degrading language-modeling quality, improving perplexity over dense RoPE at the training context length. Under length extrapolation, MoSAR achieves the best perplexity among all evaluated variants, including strong baselines such as ALiBi. Moreover, the learned geometry remains stable under deterministic top-1 discretization, suggesting that it is not only adaptive, but also amenable to low-cost approximation at inference time.

Abstract: 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.

Abstract: Web agents are usually evaluated in live environments, where environment state and judge models drift between runs, so the same checkpoint rarely reproduces the same score, making controlled studies of training phenomena impractical. We present WebMRE, an offline benchmark of 541 tasks and 5,293 steps derived from successful WebArena trajectories, with fully audited test labels and a deterministic protocol that scores a checkpoint identically on every run without any environment. Each step pairs a human oriented guide sentence with a grounded action, enabling the first study of the mutual reinforcement effect between them in web agents. Averaged over three seeds the effect holds for both models in both decoding orders and grows with scale: jointly decoding a guide lifts element selection over an action only reference by 0.9 and 0.2 points for Qwen3.5-4B and by 1.7 and 2.2 points for Qwen3.5-9B. A mediation analysis shows that the guide is a causal channel rather than commentary: forcing the gold guide as a decoding prefix lifts action accuracy from .422 to .684, another step's guide collapses it to .055, and a paraphrase that renames the target still recovers half of the gain, so the channel carries instruction meaning and not only the label string. The same channel yields an offline reward that only a replayable protocol makes computable, though optimizing it from a strong checkpoint brings no gain yet. Our fine tuned models outperform GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash, run zero shot, on every offline metric.

Abstract: Diffusion Large Language Models (dLLMs) have emerged as an efficient alternative to autoregressive models, yet aligning them via Reinforcement Learning (RL) requires likelihood surrogates estimated from masked reconstruction subproblems under a small Monte Carlo budget per rollout. Existing methods construct these subproblems by uniform random masking, leaving open the question of which subproblems to prioritize. We identify a systematic upstream/downstream structure in dLLM rollouts. Some tokens, when revealed, trigger large confidence changes in nearby undecoded positions; we call them upstream. Others induce only small local changes and are therefore downstream. We find masking downstream tokens yields substantially better-posed subproblems than masking upstream tokens, a phenomenon we term subproblem difficulty asymmetry. Based on the observation, we propose Informed Masking (IM), which derives a per-token priority score from the denoising trajectory at zero extra inference cost and biases mask sampling toward downstream tokens. IM is plug-and-play: when plugged into three state-of-the-art dLLM RL methods on LLaDA-8B-Instruct, it delivers up to 2.01%, 8.68%, and 5.77% relative average gains on math and planning benchmarks with improved training stability.

Abstract: When a language model reads an operation such as "Swap the contents of Box F and Box B", its forward pass writes keys and values for those tokens into the KV cache. Prior work on entity tracking establishes what models use: bindings are resolved at query time rather than stored as explicit latent state. We ask what they write at the operation span and how it is accessed. We split a forward pass into a frozen writer and a reader: the writer's cache is recomputed without gradients, while the reader sees only the instruction and operation tokens, with all state descriptions hidden, and is trained in isolation. Anything the reader recovers was therefore already present in the unmodified cache. On a synthetic boxes task, a base reader recovers $\leq 0.06$ of queried bindings against $0.75$--$1.00$ after training, and recoverability tracks the operation's read/write footprint. We find two modes of access. Across Llama-3.1-8B and Mistral-7B, operation-span transplants causally redirect which visible state is read even when the two worlds hold identical values, revealing a routing record. Isolation training preserves routing and adds direct access to the payload, the value the operation read, from the single operand-name token in a narrow mid-depth band (layers 12--15 of 32 in Llama-3.1-8B, 14--17 in Mistral-7B) --- the same site that holds the routing record. The same recipe extends to further operations, ToMi and GSM8K, but is bounded by training coverage and costs open-book accuracy. Operation tokens thus leave localized, causally recoverable records that support both routing and direct payload access, though the model that writes them reads mainly the address they carry and not the value.

Abstract: Looped language models repeatedly apply a weight-shared block to increase effective depth without increasing parameter count, but the resulting T sequential recurrent-block calls per generated token substantially increase decoding latency. To address the issue, we introduce Wavefront Decoding (WFD), a training-free self-speculative decoding framework designed for looped language models. WFD exploits two properties of these architectures: intermediate recurrence outputs provide effective draft predictions, and weight sharing allows token states at different positions and recurrence depths to be processed in one batched recurrent-block call. WFD organizes these mixed-depth states into a diagonal wavefront, continuously drafting new positions at shallow depth while advancing earlier positions toward full-depth verification. Unlike the phase-separated draft-then-verify schedule, WFD therefore co-batches drafting and verification within the same recurrent calls, while rejected drafts are corrected using full-depth predictions. Across six Spec-Bench task categories, WFD achieves 2.42x speedup on Ouro-2.6B and 3.54x on Huginn-3.5B over autoregressive decoding, consistently outperforming draft-then-verify. Cross-recurrence KV sharing further reduces wavefront KV traffic and increases WFD's speedup to 4.81x on Huginn-3.5B.

Abstract: Long-context inference has become a defining capability of large language models, but exact dense attention remains costly due to its quadratic scaling with sequence length. Block-sparse attention offers a hardware-friendly alternative by routing each query block to a small set of relevant key blocks, yet accurate training-free block routing remains difficult. Existing routers often pool post-RoPE token representations, which entangles semantic aggregation with RoPE-induced geometry and attenuates local positional cues through high-frequency phase cancellation. To resolve this mismatch, we propose \textbf{Semantic-Geometric Decoupled Routing}, a training-free block routing framework that shifts semantic aggregation to the pre-RoPE space and reconstructs geometric bias with an offline structural prior and relative block distances. This decomposition yields an explicit closed-form block routing score without token-level search or post-hoc calibration. Experiments on long-context text and video tasks show that our method approaches full-attention accuracy across 4K--128K contexts, keeps routing overhead below 3.4 ms, and achieves a 5.03$\times$ speedup over FlashAttn at a 128K context length.

Abstract: Steering with SAE features requires per-feature coefficient tuning, which currently demands intervention sweeps. We ask whether properties of the SAE itself, computable before any forward pass, predict which features will be cheap or expensive to steer. We show that variation in SAE feature steerability is partially predicted by decoder-space geometry: neighbor density and maximum cosine similarity to nearby decoder directions, both computable from the SAE weight matrix before any intervention, rank features by how much steering they require for a fixed behavioral effect ($ρ$ up to $-0.546$, $p < 10^{-6}$, AUROC 0.610-0.822 across conditions; the signal is rank-based, consistent with grid discreteness). This geometry-steerability relationship replicates across two Gemma-2 model scales (2B and 9B), two SAE widths (16K and 65K), and is detectable cross-architecturally on Llama-3.1-8B-Instruct ($ρ= -0.266$, $n = 300$). On Qwen3-8B with BatchTopK SAEs, geometry predicts whether a feature is steerable at all but not the continuous ordering among responsive features, revealing a boundary condition tied to SAE training regime. The signal weakens at deep proportional layer depth in both models, where the cost of steering exceeds our intervention budget, a consistent depth boundary. These results provide preliminary evidence that pre-steering geometry can partially inform coefficient selection, offering a path toward screening features for controllability before deployment.

Abstract: Memory systems for large language models have focused predominantly on efficient retrieval, whereas the decision of whether retrieved memories should be trusted has received comparatively little attention. When the memory store contains conflicting positions, standard retrieval-augmented generation (RAG) blindly injects memories and amplifies hallucinations: in models susceptible to memory injection, the RAG hallucination rate under conflicting memories is markedly higher than that of a memory-free baseline. Inspired by memory signaling mechanisms in the prefrontal cortex, we propose the Memory Decision Layer (MDL), a zero-parameter memory decision controller situated between the retrieval and generation stages. Its core is a three-signal complementary encoder that fuses relevance, reliability, and task risk through QR-based orthogonal subspace projection and a meta-working-memory signal into an interpretable decision representation that quantifies the trustworthiness of retrieved memories. Building on this encoder, MDL explicitly decouples confidence from consistency and introduces risk inversion and explicit abstention. Evaluations on mainstream large language models and multiple open-source datasets show that MDL reduces the hallucination rate under conflicting memories by about 56.04% in general scenarios and approaches zero hallucination in high-risk scenarios. The controller is fully white-box: it relies purely on geometric operations, requires no trained parameters, and adds only about 0.14 ms per decision -- roughly 50x faster than the embedding-retrieval step that precedes it and four to five orders of magnitude faster than an LLM self-evaluation call.

Abstract: Gating the value pathway of attention reportedly improves language model pretraining, and prior studies disagree on why. We argue and provide experimental evidence that such gates supply two different things that softmax attention lacks: abstention and noise filtering. The first is abstention, which allows an attention head to output nothing, bypassing the requirement that attention weights must sum to one. The second is noise filtering, which allows the value pathway of an attention head to suppress interference from superposed features in the residual stream. In our experiments in matched models from 10M to 350M parameters, we supply abstention through a learned per-head sink logit in the softmax and noise filtering through a gate on each value. We report three empirical findings. First, the benefit of abstention, measured as the reduction in validation loss relative to a matched baseline, declines as models grow, whereas the benefit of noise filtering increases with scale. In particular, abstention accounts for nearly all of the gain from gating at 10M and filtering for most of it at 350M. Second, the best model at every scale is the one with both primitives built in. Third, injecting controlled interference into the values a head reads confirms that the gate removes such interference, and reveals that each of the two gate forms we study has a characteristic blind spot. Supplying both primitives adds negligible parameters and remains compatible with the key-value cache.

Abstract: Recurrent Transformers increase computational depth through temporal recurrence, feeding each token's high-level hidden state into the computation of the next. This raises a natural question: is additional computation better spent on more temporal steps or greater physical depth? We investigate this question using Latent Recurrent Transformers (LRTs), which retain one backbone forward pass per vocabulary token during decoding and provide a controlled setting for comparing these two ways of adding computation. Specifically, we insert a latent thought token between consecutive vocabulary tokens. Each thought token passes through the same $L$ layers as a vocabulary token, sharing the backbone parameters and providing an additional stage of hidden-state refinement before predicting the next token. We compare this $L$-layer LRT against a $2L$-layer LRT without thought tokens. Both execute $2L$ Transformer blocks per vocabulary token during decoding, but the thought-token model uses fewer parameters. On 16- and 20-layer mixture-of-experts NanoChat backbones, one thought token brings the shallower model within 0.006 and 0.004 bits per byte of its double-depth counterpart, recovering 67% and 81% of the improvement with approximately 48% fewer total parameters. These results suggest that temporal thinking offers a parameter-efficient alternative to increasing physical depth in recurrent Transformers.

Abstract: Reasoning and agentic workloads increasingly demand efficient long-context inference. Yet full-attention decoding reads the growing history at every step, regardless of its benefit to the next prediction. We show that a pretrained model's decoding states already contain information predictive of this benefit, before the global read. Building on this finding, we introduce On-Demand Attention (ODA), a local-first decoding method that uses a lightweight recall head to selectively invoke global attention as its predicted benefit changes during generation. ODA trains only the recall head, leaving pretrained weights unchanged and the complete historical KV cache available for future recall. We further implement GPU-side conditional execution in vLLM, translating reduced global reads into practical decoding speedups over full attention at long context lengths. Experiments across Qwen and Gemma models, including hybrid-attention backbones, show that selective recall recovers most of the performance lost under local attention while substantially reducing global reads. These findings support long-context inference in which pretrained models guide their own access to the information they retain.

Abstract: A popular selling point of diffusion large language models (dLLMs) is their capacity for parallelism: the ability to generate sequences of text far more efficiently than autoregressive models, which require one forward pass per token. Yet among the many competing paradigms for dLLMs, from masked to uniform to Gaussian diffusion, principled understanding of how these different proposals compare in parallelism remains limited. In this work, we initiate a fine-grained comparison of the capacity for parallelism among these three leading approaches and prove the following: - Uniform and Gaussian diffusion can sample in a number of forward passes which scales with the dual total correlation of the underlying distribution, a measure of intrinsic complexity which can be much smaller than the context length. Previously, it was only known how to achieve this using masked diffusion. - For a certain family of random empirical measures, we show that $\widetildeΘ(\sqrt{d})$ forward passes are necessary and sufficient to sample using uniform or Gaussian diffusion, yet there exist approximate score oracles for which $\widetildeΩ(d)$ forward passes are needed for masked diffusion. This establishes the first provable separation in parallelism between the three prevailing dLLM paradigms. Contrary to popular intuition that masked diffusions are harder to parallelize because they must commit to token values, the latter separation instead comes from the fact that the critical windows in masked diffusion sampling are asymptotically narrower than those in uniform and Gaussian diffusion sampling.

Abstract: Two dominant tokenisation algorithms are used by modern language models: byte-pair encoding (BPE) and UnigramLM. These differ along two orthogonal axes: their optimisation objective (compression vs. log-likelihood) and their search procedure (bottom-up merging vs. top-down pruning). Existing comparisons confound these axes, making it unclear whether their observed differences stem from what is being optimised vs. how it is being optimised. We disentangle the two by introducing two new tokenisation algorithms that complete this 2x2 design space: BottomUpLL, a bottom-up likelihood-based tokeniser, and TopDownComp, a top-down compression-based tokeniser. We train language models with tokenisers produced by each algorithm, varying: model size, vocabulary sizes, and domain (English-only vs. multilingual). Evaluating models on bits-per-byte, we find that the search procedure -- not the objective -- is the dominant factor: bottom-up tokenisers consistently achieve lower bits-per-byte in most settings. Evaluating models on the BLiMP task, however, shows no consistent relationship between design choice and performance. Overall, our results disentangle the effect of tokeniser design choices on language modelling performance, offering concrete guidance for their more principled construction.

Abstract: Large language models (LLMs) trained on next-token prediction exhibit remarkable in-context learning (ICL) abilities, yet the representations that support ICL remain poorly understood. We consider such representations in a controlled setting: prompting LLMs with data emitted from hidden Markov models (HMMs) and probing for the corresponding belief state -- the posterior distribution over the HMM's hidden states given the observed token history. Across six open-source LLMs prompted with data from 40 HMMs selected for non-trivial belief structure, we find that belief states are linearly decodable from residual stream activations, with peak probe $R^2$-values from 0.83-0.99 across HMM and LLM combinations, ranging from early to late layers. To establish functional relevance, we intervene directly on the probe-identified subspace via patching and steering, resulting in downstream prediction quality on the order of the untampered model, while controls degrade performance substantially. Together, these results provide representation-level evidence that ICL in open-source LLMs approximates optimal Bayesian prediction over a context-inferred generative model. More broadly, our findings extend prior results linking input-distribution structure to activation geometry: from toy networks trained explicitly on HMM data to production-scale LLMs.

Abstract: We introduce a simple architectural modification to decoder-only transformers: a persistent recurrent state that observes hidden representations via cross-attention, updates itself through a GRU, and modulates subsequent processing via gated addition. Inserted between the lower and upper halves of a 6-layer transformer, this module adds only 3.7\% additional parameters while reducing evaluation loss from $2.438 \pm 0.004$ to $1.743 \pm 0.018$, corresponding to a 28.5\% reduction on held-out language modeling data. The improvement is statistically significant across 5 random seeds ($p < 0.01$) and corresponds to reduced overfitting (generalization gap 0.12 vs 0.26). Through controlled ablations, we demonstrate that the improvement stems entirely from the persistent memory topology, not from auxiliary self-prediction objectives. A model with identical topology but no auxiliary loss performs equivalently, while a random auxiliary loss provides no benefit. Representation probing reveals that the persistent state encodes narrative position (52\% vs 33\% chance level)---information that standard attention maintains less efficiently. Our results suggest that bridging transformer layers with a lightweight recurrent memory is a simple, effective approach to improving generalization in small-scale language models.

Abstract: KV-cache eviction perturbs the conditional token distributions governing autoregressive generation. We investigate how first-divergence timing and subsequent token mismatch determine cumulative disagreement. We derive an exact decomposition under a specified stepwise maximal coupling: the expected mismatch fraction equals a first-mismatch contribution plus post-divergence exposure multiplied by its mismatch rate. An explicit construction over unrestricted autoregressive kernel pairs realizes the sharp interval of risks compatible with a finite divergence-aligned observation window. Residual-branch conditional Monte Carlo provides unbiased joint estimates of occurrence, occupation, and window/tail contributions, with per-replicate variance dominance for total token loss. Complete trajectories from Meta-Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct show that SnapKV at 50% retention enters divergence later and less often than SnapKV-512 or recent-token retention with the same 50% prompt-cache budget, while post-divergence total variation (TV) remains high. In an exploratory analysis of 288 documents, post-divergence exposure accounts for 85-90% of four aggregate mismatch gaps. On 288 independent documents at 90% retention, prespecified comparisons show higher branch-aligned TV in the late than in the early window in both models.

Abstract: Continued pretraining (CPT) with data augmentation such as paraphrasing can store inside a large language model (LLM) the knowledge of a small source corpus. The stored knowledge, however, is not always retrieved correctly. We study the eliciting side rather than the storing side: we use preference optimization, which learns from pairs of a preferred (chosen) and a dispreferred (rejected) response, so that the model elicits its stored knowledge more accurately. One proposed approach takes the model's own erroneous response as rejected and the gold answer as chosen, so as to suppress the error. When the target knowledge is partially known, however, most of these rejected responses are factually correct. Using direct preference optimization (DPO) then pushes down rejected responses that contain correct knowledge and differ from the chosen answer only in style, such as length and wording. We propose style-debiased DPO (SD-DPO), which scores whether the rejected response of each pair is factually correct, inverts the preference of such pairs, and weights them so that the learning signal due to differences in style cancels out as a whole. We first test whether, on top of EntiGraph, a representative storing-side method that runs CPT on text synthesized from the corpus, our method adds accuracy efficiently. On QuALITY, the reading-comprehension QA benchmark on which EntiGraph was evaluated, SD-DPO exceeds a baseline we CPT on EntiGraph's synthetic data from the same base model and evaluate with the same procedure. The training tokens this requires are a few dozen times fewer than the additional CPT needed for the same gain. For knowledge updating, the main goal of this work, we use AToKE, a knowledge-editing benchmark for facts that change over time. There, SD-DPO reaches an overall accuracy of 0.982 and answers with the new or the old fact according to the queried period.

Abstract: A language model's representation geometry is not predetermined; it evolves as the model runs. A faithful account of that geometry must capture that dynamic process, and so cannot be based solely on model-independent statistics such as co-occurrence. Here we introduce a mean-field analysis of attention. The average attention from one token to another defines a kernel that carries representations layer to layer and can be iterated through the network to model how the geometry is transformed. We condition this average two ways. Conditioned on a whole corpus, the kernel predicts the average-case evolution of representation geometry. Conditioned instead on a single context, it predicts the expected geometry for that context. A head's departure from that prediction, its \emph{mean-field deviation}, isolates the context-specific computation that the mean field misses. Under the corpus-conditional reading, the kernel yields an open-loop model: from the input embeddings and the frozen weights alone, we can iterate the kernel and the model's own MLPs over token representations, never consulting a measured deviation at any layer. The resulting prediction is highly accurate. In early training the model and its corpus mean field are indistinguishable. Replace every attention head with its mean field, and the substitution leaves the loss on real text unchanged. Around the onset of induction, the two diverge, and the gap widens as representations become contextualized. Under the context-conditional reading, deviation from the mean field is a task-agnostic measure of context-specific computation. The residual decomposes additively into unusual attention routing and contextualization of the transported values. Across controlled induction and few-shot settings, greater deviation tracks greater reliance on in-context information.

Abstract: Orthrus is a hybrid autoregressive-diffusion architecture that accelerates autoregressive language-model inference by generating multiple tokens in parallel while using a frozen autoregressive backbone. Its central claim is that an intra-model consensus mechanism enables lossless speculative decoding, producing the same output sequence as the autoregressive model. We independently reproduce Orthrus and examine this claim under different numerical precisions. Under BF16 inference, exact trajectory matching occurs in only 45% of cases for the authors' checkpoint and 43% for our independently trained model across 1,190 prompts from 12 domains. The probability of exact matching is also strongly associated with the response-conditional perplexity of the reference model. Despite this trajectory divergence, Orthrus does not show systematic degradation on downstream lm-eval-harness benchmarks. In contrast, repeating the trajectory evaluation with FP32 yields exact trajectory matching on all evaluated prompts. These results show that the practical losslessness of Orthrus depends on numerical precision and that exact trajectory equivalence should be evaluated separately from downstream task performance.

Abstract: On-policy self-distillation (OPSD) scores student-generated prefixes with a solution-conditioned self-teacher, yet transfers supervision only through next-token probabilities. We ask whether the aligned final-layer discrepancy offers a useful second channel, and how to test that channel without confusing its geometry with auxiliary strength. SCOPE-OPSD projects the privileged teacher-student residual onto a frozen rank-64 factor estimated from residual covariance and language-model-head Fisher sensitivity. It reuses the forwards already required by OPSD and adds neither rollouts nor inference-time modules. A matched Random control preserves the structured factor's rank and nonzero spectrum and uses per-arm gradient-RMS calibration, isolating the effect of the data-dependent orientation. Across the complete 25/50/75/100-step trajectories for Qwen3-1.7B, 4B, and 8B, Structured is never below Pure OPSD, with strict gains in 11 of the 12 model-checkpoint combinations and an exact tie at 4B step 25. Structured also exceeds matched Random in 10 of the 12 combinations. At step 75 on Qwen3-1.7B, Structured exceeds matched Random by 1.39 Macro Avg@12 points in each of two independent training reruns. A cross-fitted diagnostic also shows 4.40 times greater held-out privileged-gap capture than the matched random orientation. The results support a compact, Fisher-conditioned privileged subspace for short-budget OPSD.

Abstract: Factual hallucination in closed-book question answering is often treated as a coverage problem: a model fails because the relevant fact is absent from its internal memory. This view misses a second source of error. Even when a fact has been observed, finite memory may force it to be stored only approximately. We study this effect through a simple coverage--compression model of factual recall. We consider an unstructured question-answering task with $N$ possible queries and $K$ possible answers. A learner observes $M$ training facts, compresses them into at most $B$ bits, and answers uniformly drawn test queries without retrieval. For a uniformly random ground-truth mapping, we prove $\mathcal{E} \geq \frac{M}{N}δ^\star\!\left(\frac{B}{M}\right) + \left(1-\frac{M}{N}\right)\left(1-\frac{1}{K}\right)$, where $δ^\star(r)$ is the inverse rate-distortion function of a uniform $K$-ary source under zero-one loss. The two terms separate compression distortion on observed facts from missing coverage on unobserved facts. The bound gives a compact way to reason about selective memory, forced compression, structure, retrieval, abstention, and long-context organization. We study the predicted signatures with theory-implied simulations and controlled fact-injection probes in modern language models that vary fact load and effective trainable memory. The result is not a complete theory of hallucination, but an information-theoretic account of a separable failure mode: lossy recall of observed facts under finite memory.

Abstract: How does a language model's dependence on query-routing information and target knowledge change as it answers a question? We study this question through layerwise interventions on the hidden state at the end of the question. Across Qwen, Llama, and Gemma, we compare country-continent questions with noun, adjective, and code answers while keeping several fitted measurements distinct. A pair-conditioned request direction describes which country is queried in natural single-country questions; a global request direction describes first- versus second-country requests in paired questions; separate selection candidates test control among contents already available in the hidden state. A diagnostic reanalysis of frozen Qwen natural-question states shows that the pair-conditioned direction grows stronger before interventions on it begin to alter later fitted knowledge, with this causal window opening while answer-supporting content is still forming. The paired three-model trajectories are not uniform: Gemma shows a partially overlapping mid-layer routing-content profile, whereas Llama has no sustained routing-effect window under the same gates. In the paired protocol, dependence on the global request direction decreases from fixed earlier to later layer sets while dependence on fitted content persists. A matched Qwen comparison shows that the pair-conditioned direction retains a late effect, so this operational handoff concerns the global fitted direction rather than all request information. These results separate early readability, natural strength, causal steering, and later content dependence.

Abstract: Self-supervised speech foundation models are now used in a wide array of downstream applications, including traditional speech recognition and as the basis for tokens in speech-aware language models. Attempts to understand their usefulness have largely focused on probing their representations' ability to discriminate phonemes and words. However, discriminative ability for words need not imply specialized representation of words per se. Good discrimination of words may be explained by good encoding of word form (phonemes) rather than form-independent word representations encoding identity or syntactic/semantic properties. By partialling out phoneme information using residualization, we show that, in later layers, HuBERT and wav2vec 2.0 do in general learn representations which encode words with reasonable fidelity independently of local phonetic content. We show that this simple approach to disentanglement can enhance higher-order linguistic information in word discovery tasks.

Abstract: Reading a transformer's internal states in token space is easy to do and hard to trust: a logit lens on a single hidden state is dominated, at intermediate layers, by the generic tokens the model would predict for almost any input. We read the difference instead. Subtracting two closely matched prompts' hidden states and projecting through the unembedding cancels the shared component and surfaces what separates them, an operation equivalent to reading a RepE/ActAdd steering vector through a logit lens. Built into a training-free tracer that reads at every position, sub-layer, and head and averages over designed baselines, it traces a compound- noun MLP->attention chain in Phi-2, confirmed there by activation patching, with the same distinction recovered across three architectures by readout and probe rather than by patching; it reads what retrieval surfaces for real versus fictional entities, and reads metaphor as a set of domain-to-domain mappings rather than a single figurativity feature. A cross-seed control marks the boundary: across five networks differing only in initialization, the same distinction surfaces as almost entirely different tokens (top-10 overlap 0.08). What a computation looks like in token space is network-specific; the distinction it draws is not

Abstract: Long context language models now advertise windows of one million tokens, but two habits limit how much of that window is used. Attention heads with nothing useful to read still spend their budget on the first token, which is called the attention sink, and where a fact sits in the context changes whether the model finds it. Gated attention cut first token attention from 46.7 percent to 4.8 percent at NeurIPS 2025, and Kimi K3 pairs that idea with Kimi Delta Attention and Attention Residuals behind a one million token window, eight times past the range where these diagnostics have been reported. This paper asks whether the fix survives that jump. We build SinkProbe, a suite that measures sink mass, massive activation, position resolved recall and the recency gap, and apply it to four small models that differ only in how they mix tokens and depth. Three results follow. The training objective produces the sink, not the architecture. Gating did not reproduce its published effect at our scale. Sink mass, activations and position bias moved independently. Code, data and the measurement protocol are released at https://github.com/sararizwan7/Attention-Mechanisms-in-1M-Context-Window

Abstract: On-Policy Distillation (OPD) facilitates the transfer of knowledge from domain expert to student in the post-training phase of Large Language Models (LLMs). However, the supervision signals in mainstream OPD methods suffer from high variance and noise which is generally instable during training. In this work, we systematically investigated what really matters to the performance and the fundamental mechanisms behind the instability during training. We found that retaining only the sign of token-level advantages is sufficient to achieve the performance comparable to standard OPD. Meanwhile, smoother and bounded advantages can stabilize the training process without sacrificing its performance. These motivated us to shape the advantages using the Total Variation (TV) and propose a robust TV regulated On-Policy Distillation (TV-OPD) method. Benefiting from the bounded and diminished advantages, TV-OPD exhibits stable training dynamics and steady late-stage performance. We conducted comprehensive experiments and found that, across various settings, TV-OPD consistently achieved better performance and lower variance in the late-stage of training.

Abstract: Linear attention is increasingly used in frontier language models for efficient long-context inference and constant-memory decoding. Its fixed-size recurrent memory, however, requires an online decision at each token: what to write and how strongly to overwrite existing associations before knowing which information future queries will require. Delta-rule models learn this strength from the current token embedding but do not track confidence in the memory estimate, preventing each write from adapting to accumulated evidence. To represent this uncertainty explicitly, we reformulate recurrent associative memory as a linear--Gaussian state-space model, for which the Kalman filter is the optimal recursive estimator, and introduce a new family of models, Kalman Delta Networks (KDNs). Within KDNs, the transition propagates both the memory state and its uncertainty, allowing the Kalman gain to weight each residual write by accumulated evidence and observation reliability. Under this formulation, Delta-style updates emerge as a special case that substitutes a token-wise isotropic surrogate for predictive covariance and omits covariance tracking. Exact tracking, however, entails a dense, state-dependent Riccati recursion that is poorly suited to GPU-parallel linear-attention scans. To address this issue, we introduce two scan-compatible KDN approximations. Diagonal KDN projects each one-step posterior onto the diagonal Gaussian family through online mean-field variational inference, whereas Isotropic KDN uses an isotropic approximation with a single uncertainty scalar per head. Their uncertainty recurrences are Mobius maps, enabling associative scans with logarithmic parallel depth. Across controlled pretraining at 750M and 1.3B parameters, KDN variants consistently improve perplexity and mean downstream accuracy over state-of-the-art linear-attention models.

Abstract: Methods for streaming language models are often discussed alongside long-context and memory systems, although they solve different problems. An attention sink can stabilize autoregressive generation over an indefinitely long stream while the model remains unable to use content that has left its recent-token cache. We argue that this distinction should be explicit in system claims and evaluation. We introduce three horizons: the stability horizon, over which predictive behavior remains well behaved; the access horizon, over which past content can still causally affect the output; and the utility horizon, over which a task retains acceptable performance. We show constructively that the stability horizon can be infinite while the access and utility horizons are finite. We then propose ThreeH, an evaluation contract that measures all three horizons under a common state and compute budget. Applying the framework to attention-sink streaming clarifies its strength, constant-memory, stable generation, without treating anchor tokens as semantic memory. The framework exposes roles for cache policies, recurrent state, retrieval, and external memory. Experiments on 128K-token streams, delayed binding recall, and delayed decisions show that attention sinks preserve local modeling but not content beyond the active cache; recurrent and retrieval state extend the semantic horizon.

Abstract: Autoregressive language models generate one token per decoding step, limiting the useful output of each forward pass. Although diffusion models, insertion-based decoding, and multi-token prediction enable parallel generation, they either incur additional training-time token traffic or struggle to predict strongly dependent future tokens. We introduce the Line-Coupled Language Model (LCLM), an autoregressive model that advances multiple text lines together by predicting the next token for every active line while coupling the lines through shared causal context. LCLM interleaves line tokens into a single causal sequence and uses line-staggered rotary positions, retaining the standard next-token objective and causal attention. Controlled experiments show that cross-line targets are substantially less dependent than consecutive same-line targets, supporting lines as parallel generation units. With 881M parameters, LCLM produces an average of 2.94 content tokens per forward pass with a validation cross-entropy loss of 2.44, compared with 1.00 token per forward pass and a loss of 2.39 for the vanilla autoregressive baseline. Most notably, even when LCLM generates 16 tokens per forward pass, its loss is only 0.09 higher than that of the vanilla autoregressive baseline (2.34 vs. 2.25).