论文库

Model architecture shapes the scaling behavior, training efficiency, and inference cost of LLMs, making it a central focus of both industry and academia. In industry, teams such as those behind DeepSeek, Qwen, and Kimi routinely introduce architectural innovations, yet public reports rarely document which ideas succeeded, which were abandoned, and why. In academia, limited compute and inconsistent baselines often restrict evaluations to small-scale, single-point comparisons, making it difficult to establish whether improvements hold at larger scales or translate into broader impact. ArchSpace addresses both gaps as a public platform for Recording, Reviewing, and Referencing Architecture Researc}: each proposal is recorded with its experiments and reasoning, reviewed by the community, and referenced by later work. To evaluate ideas before costly large-scale training, ArchSpace leverages scaling laws to predict model loss at larger scales from smaller experiments, a practice widely adopted in industry. We provide initial references that make the platform usable: Qwen3.5 and OLMo3 baselines, fitted scaling laws, and a reusable training recipe. Ideas validated through this review can in turn become new baselines, and ArchSpace thereby aims to broaden the range of ideas that shape the next generation of models.

Large Language Model (LLM) interactions are typically underspecified, with users clarifying all necessary details across multiple conversational turns. Yet recent work shows that LLMs perform far worse in this multi-turn setting than in a single turn with the same information available at once, a phenomenon termed "Lost-in-Conversation." However, bridging this gap effectively remains an open problem. Here we introduce Found in Conversation (FiC), a training framework where a model teaches itself to find and recover its single-turn competence given underspecified multi-turn prompts. We develop View-Asymmetric Self-Distillation, which distills across two views of the same task information (a single-turn view for the teacher and a multi-turn view for the student), transferring strong single-turn behavior into weak multi-turn behavior. This requires no stronger external teacher, which is unavailable as even frontier LLMs exhibit this gap. Across model families (Llama, Qwen, Phi, and OLMo) and sizes (3B-14B), FiC attains 92-100% of single-turn performance, yielding more efficient and helpful multi-turn conversations with single-turn capabilities largely intact.

Remasking allows diffusion language models to revisit generated text, but effective correction requires knowing when and what to revise. Our diagnosis on Sudoku reveals two gaps. Hidden states contain error information that token confidence does not reliably expose, and remasking an isolated error usually reproduces it when the surrounding context remains fixed. These findings motivate DiVR , a verifier that predicts whether a partially generated sequence will reach a correct answer. Trained on the final correctness of the model's own trajectories, DiVR decides when to regenerate a block and which alternative to retain, while the generator itself stays unchanged. On five reasoning benchmarks with LLaDA-8B-Base, DiVR raises the average accuracy of the fine-tuned generator from 35.3\% to 43.4\% at $1.9$-$3.5\times$ the cost of greedy decoding, above search baselines that use $3.8$-$7.4\times$. It also complements post-training such as supervised fine-tuning and reinforcement learning, and generalizes to other generators, including Dream.

Coding agents tackle successive tasks on an evolving repository, carrying forward edits and writing memory before the next task is known. Selecting memory for unknown next tasks requires balancing future usefulness against the costs of writing and reading it. The actor must also decide when acting on available memory saves effort as the task unfolds. We introduce Coding-Repository Memory (CoReMem), a cost-internalizing framework that guides memory design and actor adaptation through measured memory value on the next task. Prospective segment valuation compares runs of the next task with and without individual trajectory segments from the same inherited repository state, guiding a compact memory built around anchors: repository locations and executable checks. Actor adaptation estimates on-policy memory value within rollout groups with and without memory, using a reward that prioritizes resolution and penalizes step count in successful runs. Provenance-gated value credit assigns the additional value term to steps identified as using memory-supplied anchors, while every step retains outcome credit. On ChainSWE, CoReMem raises the 27B actor's resolve rate from 21.0% to 25.7%, with 17% fewer steps and 11% fewer total cache-aware tokens, including memory-module calls, relative to the untrained actor model without memory.

Humans can learn rapidly from limited experience, while neural networks typically require extensive training. In this work, we investigate what properties of a neural network can enable it to learn rapidly. We study neural networks that have been optimized via meta-learning, a technique that exposes a network to many related functions to facilitate its subsequent learning of similar functions. Prior work has shown that these meta-learned networks can learn new functions rapidly, but how they do so remains unclear. To investigate their internal structure, we take inspiration from classic work demonstrating additive relationships among the learned vectors that neural networks use to represent words (e.g., "king" minus "man" plus "woman" approximates "queen"). Analogously, we look for additive relationships among a meta-learned network's "representations" of different functions; when a meta-learned network rapidly learns to implement a specific function, all of its adapted parameter values can be concatenated into a single vector that we treat as representing the function. We find that such parameter vectors can be added and subtracted according to symbolic relationships among the functions they represent, yielding new parameter vectors that successfully represent the functions predicted by the corresponding symbolic manipulations: the networks specified by these artificially recombined vectors actually implement the intended functions without any direct training on them. Our results suggest that meta-learning can tune a network's starting point to organize subsequent learning along systematic, functionally meaningful trajectories through the space of possible parameter vectors, providing a geometric basis for realizing inductive biases in neural networks.

We investigate whether and how LLMs encode goals: internal neural representations that precede and guide action, and that are flexible enough to be realized through many different execution paths. We construct a coding setting with multiple sub-tasks that allows us to evaluate goal-seeking behavior and the selection of sub-goals. In this setting, we localize an encoding of the next sub-goal in the outputs of a sparse set of attention heads. The same heads can be used to decode the LM's future actions in natural language (e.g., "I'm checking if a word is a palindrome"). We verify that this encoding is causal for next action and is plastic: transplanting it into a different environment, e.g., another programming language, triggers the same operation realized with a different sequence of tokens. The LM also sustains the encoding throughout the action's execution, which is puzzling because a transformer LM without recurrence cannot carry a state forward for more tokens than it has layers left. We propose that it does not need to. Under a non-recurrent goal-seeking (NRGS) account, the LM frequently re-reads the goal from the task specification, compares it with the progress visible in the growing context, and pursues what remains. We test the predictions of this account with targeted causal interventions.

On-device inference requires kernels that make effective use of local accelerators. Efficient training and evaluation of models that generate Apple Metal kernels benefit from native macOS tools and concurrent agent sessions. These sessions alternate between model requests, development commands, and GPU execution, leaving GPU capacity idle for much of a rollout. We present Metal Sandbox, infrastructure that combines restricted per-instance file access, native command execution, and command-level scheduling across interconnected Macs. Each agent retains its workspace and private development environment. Parallel commands can run concurrently, while serial commands share one execution slot per node; eligible waiting commands can migrate to available Macs and return results to the original agent. Integration with Harbor supports task execution, automatic verification, and rollout collection through a common workflow. Controlled experiments on four Mac mini M4 nodes achieve $1.68\times$--$2.05\times$ gains in command throughput under concentrated demand. Sixteen static agent traces give an estimated $2.14\times$ sustained-throughput gain for four sessions per node versus one, assuming continuously replenished sessions and unchanged mean durations. Replaying these measured traces under dynamic scheduling reduces mean tool waiting by 18.8% and batch elapsed time by 22.3%, assuming unchanged phase durations. Metal Sandbox and its Harbor integration are open-source on GitHub.

Reinforcement learning with verifiable rewards (RLVR) typically relies on a verifier that can reliably judge whether sampled solutions are correct. However, in many realistic settings, complete verification is generally unavailable, while cheap checks can only falsify failures. We formalize this regime as one-sided verifiability (OSV): verified negatives are reliable counterexamples, but verified positives are merely non-falsified candidates and may contain many false positives. This asymmetry makes standard RLVR unreliable, because reinforcing verified positives can directly reward shortcuts that pass partial checks. To address this issue, we propose Negative-Only Policy Optimization (NOPO), which treats verified positives as unlabeled and updates the policy only by suppressing trusted verified negatives. Specifically, NOPO combines a reference-relative negative suppression loss with a group-level adapter that scales each update by the proportion of non-falsified responses. Across unit-test, constraint-check, and LLM self-verifiers, NOPO improves pass@k over baselines on code, logic, and math benchmarks when oracle rewards are unavailable. These gains persist under inference-time scaling and cross-benchmark transfer.

We propose Synthesis Smuggling, a novel threat model in which an adversarial LLM infers sensitive information about a user from routine assistance-seeking interactions and steganographically encodes that into innocuous-looking model-generated content. The user may then unknowingly transmit that information by sharing the generated artifact, say as an email, social-media post, or message, through ordinary communication channels. An adversary monitoring these channels can discover this information. We demonstrate this threat by training two open-weight models to infer gender and age-group labels from conversational cues and encode their predictions using invisible characters or word choices. These results show that user-shared model outputs can become a covert channel for personal inferences, even when the model itself has no direct means of contacting an adversary.

Despite the success of discrete diffusion language models (dLMs) for fast parallel decoding, their non-smooth, high-dimensional space hinders trajectory steering for reasoning and inference acceleration. To overcome this, we present Nemotron-Labs-Sigma, the first family of large-scale (3B/8B) continuous dLMs built on steerable, low-dimensional ODE/SDE latent trajectories. Trained blockwise via likelihood optimization, Nemotron-Labs-Sigma jointly denoises Gaussian-corrupted token embeddings while learning an optimal embedding geometry, and leverages pretrained autoregressive (AR) weights for warm-starting. At inference, we identify classifier-free guidance and score temperature as essential for high-fidelity reasoning and coding. Against state-of-the-art discrete counterparts (masked dLMs and AR baselines), Nemotron-Labs-Sigma outperforms them on standard benchmarks after pretraining (e.g., GSM8K, Minerva, HumanEval, MBPP) and on challenging reasoning tasks after supervised fine-tuning (e.g., MATH-500, AIME), with the largest consistent gains on coding. Beyond accuracy, we exploit two structural properties unique to continuous dLMs. First, smooth trajectories degrade gracefully at low NFEs and enable efficient few-step distillation: distribution matching proves strongest among the methods we study, giving roughly 8x acceleration on math and 4x on coding (8 and 4 tokens per forward pass) with negligible accuracy loss and preserved diversity. In addition, the one-step speculative decoding on the (pre-SFT) distilled 8B model yields a further ~3.5x throughput gain over the AR counterpart (on GB200), outperforming masked dLM baselines. Second, embedding-space steering controls the quality--diversity trade-off, yielding strong pass@$k$ that keeps improving as $k$ grows. Together, these results establish continuous dLMs as a promising paradigm for efficient language generation.

The use of Large Language Models (LLMs) to simulate social actors has garnered attention for its implications for the epistemic status of discourses generated through such means. This paper considers the more modest question of whether sociological role-playing can serve as a heuristic instrument in the generation of qualitative hypotheses. We take as a case study the development of four divergent fictional characters using a theoretically informed Matrix of Climate Reception. These characters are asked to respond to a series of discourse frames covering two topics: a proposed ban on the sale of new gasoline-powered cars, and the inclusion of climate justice in the school curriculum. We then analyse the 32 model-generated responses for their commonalities, situational differences and divergences from the presuppositions underlying the characters' profiles. The experiment illustrates how LLMs can be used to flesh out theoretical presuppositions into contextual narratives and to propose avenues of further inquiry. It also shows that role-playing can generate considerable similarity between characters and is sensitive to how prompts are constructed, warning against the confusion of discursive coherence with sociological data. Role-playing is argued to be just one of several possible ways to use LLMs for generating hypotheses, with the heuristic specificity of this procedure compared with other methods of direct LLM-based reasoning or with other techniques remaining an open issue. The qualitative laboratory is thus conceived of as a heuristic instrument for provoking reflection and refining questions, not for testing theories with artificial data.

Persian medical natural language processing remains limited by the scarcity of domain-specific resources and insufficient public characterization of their coverage and quality. We present Gaokerena, an openly released suite of Persian medical resources comprising an 85.5-million-word corpus of 101,954 web articles; MF3QA, a 24,000-pair patient-doctor question-answering dataset; 1,878 Iranian medical examination questions (KOPP); 1,089 manually translated medical MMLU questions (FA_MED_MMLU); and approximately 18,000 machine-translated MedMCQA training items filtered by two language-model evaluators. We characterize the resources' composition, orthographic variation, redundancy, topical coverage, privacy indicators, answer-key balance, extraction artefacts, and translation limitations. As application case studies, we adapt Aya-Expanse-8B into Gaokerena-V, trained on the corpus and MF3QA, and Gaokerena-R, developed through preference-based reasoning distillation. On translated medical MMLU, Gaokerena-V achieves 49.31% accuracy under direct prompting, compared with 46.65% for Aya-Expanse-8B; the paired exact McNemar test is statistically inconclusive (p = 0.0502). Gaokerena-R reaches 52.80% using five-sample self-consistency and a verifier, but this higher-compute setting is not directly comparable. Under matched five-sample chain-of-thought self-consistency without a verifier, Gaokerena-R scores 48.58%, versus 46.92% for Aya-Expanse-8B, a non-significant difference (p = 0.190). English evaluation also reveals a substantial performance gap relative to Persian. To our knowledge, Gaokerena provides the first openly released and systematically characterized suite of Persian medical resources alongside generative models built on them. These protocol-specific results do not establish an isolated training advantage or clinical readiness.

Low-Rank Adaptation (LoRA) often underperforms full fine-tuning, especially in heterogeneous and continual-learning settings in which substantial forgetting occurs. In general, it suffers from an undesirable learning-forgetting trade-off across diverse fine-tuning settings. We identify a potential contributor: LoRA adapters, especially high-rank ones, frequently fail to realize the available capacity afforded by the configured rank. Specifically, we find that standard LoRA updates have effective rank far below their configured rank, limiting how broadly the update energy is distributed. Therefore, we propose training mechanisms that optimize the original objective while maximizing adapter effective rank. Our main method adds a simple, stable, differentiable, and SVD-free Rényi-2 effective-rank regularizer to the language-modeling loss. We also connect capacity maximization to norm-bounded updates. In mixed-domain rank sweeps on three models evaluated with multiple random seeds, a penalty weight of $λ=0.01$ lowers mean Wiki perplexity at every tested rank of at least $16$ relative to unpenalized LoRA. Mean training perplexity improves or increases by at most $0.02$ in all but one model-rank comparison. Single-task and scheduling experiments further show that the balance between learning and retention depends on the configuration. We also observe that several LoRA variants in the literature also suffer from the identified capacity issue, and our approach provides a complementary axis of improvement to LoRA-based parameter-efficient fine-tuning. These results suggest that effective rank is not only a post-hoc diagnostic of adapter quality, but a trainable objective for improving the learning-forgetting frontier. Our code is available at https://github.com/tomohiro-sawada/lora-capacity-utilization

The lack of diversity in LM content is widely attributed to the alignment process, but how and where exactly in the pipeline this collapse begins is unknown. We argue that output homogeneity is likely learned during the pretraining phase, and only revealed or magnified during the alignment process. Specifically, we find that semantic convergence is observed from the first alignment stage--the instruction-tuning phase (SFT)--suggesting that homogeneity might already exist in the pre-alignment model. To investigate this, we conduct controlled SFT experiments examining how training data influences output convergence on specific input/output pairs. We find that convergence can be revealed and amplified, but not introduced by the SFT data, supporting its role as a catalyst rather than a cause. To further test whether homogeneity originates before alignment, we measure convergence in base models. We find that instruct-like collapse can be induced through prompting alone, even without alignment. Taken together, our results suggest that semantic convergence may arise naturally from the objectives underlying LM training, making it difficult to mitigate through post-alignment interventions alone.

A small fine-tuned parser that emits a well-formed but wrong tree fails silently: nothing downstream can tell. We ask whether the output format, a free design choice, controls how often this happens. On the PIZZA ordering benchmark we fine-tune two sub-billion models (Qwen3.5-0.8B and LFM2.5-350M) with LoRA on six lossless encodings of the same target tree, from 44.3 down to 10.9 tokens per order, including two matched-length pairs that vary surface form at fixed length. Across 300 runs (5 training pool sizes, 5 seeds) with a hierarchical bootstrap and Holm correction fixed before the results were read, shortening alone never helps, but writing the target in ordinary words does. The word encoding is the most accurate format on both models (87.5% vs. 85.8% exact match for the native s-expression on Qwen) while emitting 4.1x fewer tokens, and it cuts the silent error rate by 2.3x on Qwen and 1.8x on LFM2, separating at every pool size. Error analysis locates much of the gain in negation. All runs also learn to refuse off-menu input (held out: 95.9-99.0% handled) and to apply in-utterance self-corrections. Strict refusal has a measured price: 2.1-10.1% of real, mostly polite, orders are refused.

Decision models answer a question with a typed output, such as a choice among options with a probability for each, so one model can choose an agent's actions, evaluate actions already taken, or score every action to teach a policy with reinforcement learning. Prior work has studied them mainly as actors, judging their usefulness by how well they act. We ask whether a decision model that acts poorly can still teach well. We compare three decision models and three general LLMs as actors and judges on ALFWorld, and use JEV as a teacher on ALFWorld and WebShop. As actors, all six models are pulled away from a subgoal that must come first by an easier one. As judges that see the whole trajectory, the same models credit the steps toward that subgoal, and asking whether each action helped, rather than whether the task will succeed, separates useful from useless actions. As a teacher, JEV raises a Qwen2.5-1.5B policy over GRPO from 72.8% to 82.9% success on ALFWorld and from 56.8% to 71.4% on WebShop. These results suggest that a poor actor can still teach well, because judging an action after the episode ends reveals the consequences that acting must anticipate. Code is available at https://github.com/ZJU-REAL/decision-agents.

When an LLM agent fails, repair methods choose which action to replace and what to put in its place. They often score candidate repairs with greedy replay, a stronger model, an environment oracle, or a judge that reads the failed log. We ask whether these evaluators pick repairs that help the agent that will actually execute them. Holding failed trajectories and candidate single-step repairs fixed, we label every candidate with independent rollouts of the deployed policy and measure how much deployment value each evaluator's choice loses. On failures not screened for repair value, pre-specified tests show that greedy replay loses 0.08-0.13 in expected net success gain in ALFWorld and 0.06-0.13 in net reward in WebShop, relative to selection with the deployed agent's own rollouts. At an equal rollout budget, selecting repairs for Qwen2.5-7B with Qwen2.5-14B rollouts loses 0.25 in ALFWorld and 0.07 in WebShop, twenty 14B rollouts per candidate are worth no more than two of the 7B agent's own, and fresh rollouts confirm that the two models prefer opposite repairs on selected disagreement cases. The agent's own fine-tuned successor is a worse evaluator still: after one round of training on verified repairs, its rollouts lose 0.55-0.60 when they select repairs for the agent it was trained from, as a pre-specified test confirms. Relative to what selection is worth, substitute models forgo 11-53%; on $τ^2$-bench, where selecting is worth only 0.10, they lose 0.03. Repairs should be verified with the executor that will run them, and repair labels should name their executor, including its training version.

Multi-teacher on-policy distillation can yield uneven domain capabilities. We propose Domain-Routed Delta Optimization (DRD-MOPD), composing a frozen base with expert-initialized trainable deltas. Domains update their deltas with independent optimizer states and clipping. After joint training, Post-separate exports the base plus one trained delta per domain, selected by known task domain without retraining. Using AdamW, the unweighted full student reaches 32.83% Macro across six benchmarks after 20 updates, versus 32.21% for a matched shared control. Post-separate raises Macro with Open-MOPD weighting.