Papers

LLMs handle ordinary typing variation fluently: a typo or missing punctuation leaves both user intent and the model's response substantively unchanged. Yet probes that detect malicious prompts by reading the model's hidden states tell a different story: the same edit rotates the readout vector by 43–56 at the perturbed token, decaying below 15% within ~10 downstream tokens. Stacking ~3 common typos per message cuts a single-position prompt-injection probe's TPR@FPR$=1% by 12.0pp, a gap recalibration alone cannot close. Multi-position aggregation cures localized perturbations (<= 0.5 loss) but only attenuates distributed ones, where even attention- and max-based aggregators still drop ~3.8pp. For single-position probes, we introduce a KV-cache fork: a short fixed suffix appended after the user message lets the probe read a few tokens downstream of the perturbation, exploiting its rapid spatial decay. This closes 95% of the gap (-0.6pp residual) – an order of magnitude better than perturbation-augmented training (-3.7pp). The rotation-and-decay geometry replicates on Llama-3.1-8B, Qwen3-8B, and Gemma-4-E4B; probe evaluation is on Llama-3.1-8B. Code: https://github.com/eladd-ai/latent-undertow

Automatic Text Summarization (ATS) in Natural Language Processing has been an important task in Information Retrieval. It compresses a document to create a summary that captures all the relevant and important information conveyed in the document. This study explores Hypergraph for extractive text summarization of single documents. Objective: This study explores a novel method of leveraging the property of domination in hypergraphs to generate an extractive summary and compare its performance with state of the art graph based methods. Method: Our work aims to generate an extractive summary by creating a sentence hypergraph where each sentence represents a node and the edge is a keyword or a named entity that contains the sentences in which it occurs. We generate a hypergraph where each edge is a keyword or an important topic and the nodes are sentences containing those keywords. Then we apply a greedy algorithm to find the dominating set of the hypergraph which will contain sentences that will form the extractive summary.

Recent advances in large language models (LLMs) have made prompts increasingly large and complex. Techniques such as chain-of-thought reasoning (Wei et al., 2022) and in-context learning (Brown et al., 2020) frequently push real-world prompts past several thousand tokens, increasing inference cost and latency. Learned compression methods such as LLMLingua (Jiang et al., 2023) and Selective Context (Li et al., 2023) achieve high compression ratios but require auxiliary language models and are non-deterministic. We ask a complementary question: how far can a training-free, fully deterministic, CPU-only pipeline based on classical lexical NLP be pushed before output quality degrades significantly? Eleven toggleable lexical transformations - stopword removal, filler-phrase deletion, contraction and abbreviation substitution, part-of-speech-based pruning, lemmatization, WordNet-driven synonym shortening, and named-entity preservation - are assembled into a configurable pipeline. Fifteen configurations are evaluated on 1,242 English-only prompts from six sources (Dolly-15k, LMSYS-Chat-1M, WildChat-1M, MMLU, GSM8K, HellaSwag), spanning eleven automatically derived task categories, yielding 18,630 paired GPT-4o-mini completions. Output preservation is measured using BLEU, ROUGE-1/2/L, BERTScore-F1, and SentenceBERT cosine similarity. The most aggressive configuration achieves a mean token reduction of 40.3% (sigma = 9.2) at a BERTScore-F1 of 0.876 against the original-prompt output; a stopword-only configuration achieves 29.6% reduction at 0.913. The compression-versus-fidelity Pareto frontier is characterized per task category, with commonsense reasoning a systematic failure mode under aggressive compression. All code, prompts, and per-cell results are released for reproducibility.

A personalized agent needs a user memory: a persistent model of who its user is. Today it is almost always text – transcripts and captions retrieved by similarity. This serves the captionable half of a person ("my cat is named Bibi"), but discards the perceptual half no caption can hold: how a voice sounds, how a face reads across age and lighting, how tired someone sounds. We measure this loss across five modalities: a strong caption-based re-identifier recovers as little as 0.11 of a dedicated encoder's recall, collapsing toward chance on non-nameable signals. We instead ground perceptual memory in the model, decomposing recall into two subproblems: a vision-language model grounds the referent in context (what and where), and a dedicated encoder extracts an identity key (who), stored as one inline token read by attention at generation with no external round-trip. Neither suffices alone – the VLM identifies cross-age faces at only 0.54 recall where a face encoder reaches 0.81, and an ungrounded encoder recognizes a two-person-scene referent at 0.05 – yet together they reach correct-region oracle (0.96), generalizing to multi-speaker audio and video. The recognition core is training-free: it reproduces the encoder's recall on any frozen model at O(1) registration cost. On PerceptMem (12 domains, 1,080 tasks) perceptual identity is capacity-limited while exact facts are binding-limited: identity belongs in a parametric bank, facts in a text store. The two memories compose cleanly: an agent with both can remember not only what its user said, but also what they are like.

Ancient Indian medical texts like Sushruta Samhita have extensive information on diseases, treatments, and surgical techniques. Yet, their ancient format and use of intricate vocabulary pose difficulties in accessibility and systematic ordering. The research here utilizes Natural Language Processing (NLP) methods like Named Entity Recognition (NER), BERTopic modeling, and Knowledge Graph development in Neo4j to extract, categorize, and visualize important concepts based on translated versions. Thematic classification with BERTopic allows for the identification of the underlying medical topics, whereas NER supports the structured entity recognition of diseases, treatments, researchers, and medicinal plants. Graphbased network analysis with Neo4j also allows for the semantic representation of relationship among extracted entities, supporting knowledge retrieval and digital preservation. The findings illustrate how graph databases, topic modeling, and entity recognition facilitate the computational organization of Ayurveda's historical medical wisdom, closing the gap between the conventional texts and contemporary data-driven inquiry. The suggested method promotes historical text analysis, medical informatics, and digital humanities to make ancient Indian medical wisdom more accessible and understandable.

Pharmaceutical sponsors developing a drug for both the United States and the European Union must reconcile guidance issued independently by the FDA and the EMA. Where the two agencies require substantively the same thing, a sponsor can file once; where they diverge, a single trial design risks rejection in one region; where one agency is silent on a point the other regulates, the sponsor must infer obligations. Today this reconciliation is performed manually by regulatory-affairs experts. We introduce cross-jurisdiction regulatory divergence detection: given an FDA requirement and an EMA requirement on the same topic, classify their relationship as AGREE, DIVERGE, or SILENT. SILENT is inherently directional (SILENT_FDA vs. SILENT_EMA); we record direction per pair and report per-direction F1 alongside the collapsed label. We release RegDivergence-101, a 101-pair expert-grounded pilot evaluation benchmark (labels grounded in three peer-reviewed FDA/EMA comparison studies and primary FDA/EMA/ICH guidance text; dual-annotation inter-annotator kappa = 0.85), and systematically characterise a four-method baseline hierarchy: lexical heuristic (0.511 macro-F1, 95% CI [0.411-0.605]), NLI cross-encoder (0.233), obligation-level Graph-RAG (0.663 [0.570-0.747]), and flat LLM judge / Claude Haiku (0.830 [0.747-0.908]). Three directional observations emerge at pilot scale (n = 101): SILENT is semantically detectable but invisible to entailment-only formulations; pair-level obligation graphs improve over lexical methods but trail flat-LLM context (CIs partially overlapping); and corpus-level graph construction is the indicated architectural target for large-scale silent-detection. RegDivergence-101 is a pilot release establishing the task formulation and baseline hierarchy; four unrepresented regulatory domains and an expansion roadmap are described in Section 7.

Recent work trains reasoning models with length penalties to curb overthinking and cut inference cost. We show that these penalties make the chain of thought less monitorable. A length-compressed model still lets misleading hints steer its answers, but it less often verbalizes their influence. We train Qwen3-4B and Qwen3-14B with reinforcement learning under length penalties targeting 60% down to 30% of baseline chain-of-thought length, then evaluate them with nine types of biasing hints on held-out MMLU-Pro-R and four transfer benchmarks. A chain is faithful when an LLM monitor can tell from it that the hint influenced the answer. At the 30% target, accuracy stays near baseline and wrong-answer hints switch answers as often as before. Yet faithfulness drops on every evaluation set for both models, by 39% for Qwen3-14B and 35% for Qwen3-4B on MMLU-Pro-R. A control trained with the same correctness and format rewards but no length penalty leaves faithfulness intact or raises it. Shortening alone does not explain the drop. Compressed chains mention the hint 7 to 35 percentage points less often than the uncompressed model's chains shortened to the same length by random sentence deletion, across both model sizes and all five evaluation sets. Length penalties therefore trade monitorability for inference cost by removing the evidence monitors depend on.

Identifying faithfulness hallucinations in LLM-generated outputs remains challenging due to the scarcity of high-quality annotated data. Recent work relies on advanced LLMs to synthesize training data, including rationales, labels, and hallucinated claims. However, these methods treat the generator as a static component, limiting iterative improvement of the detector. To address this limitation, we introduce Hallucination Self-Play (HSP), a novel framework that enables the detector to bootstrap with an evolved generator. HSP involves two roles initialized from the same base model, a detector that assesses the faithfulness of model outputs, and a generator that produces increasingly hard-to-detect hallucinated responses. Specifically, the detector is first fine-tuned on human-labeled data and then employed as a reward model to train the generator via reinforcement learning from AI feedback (RLAIF). In turn, the evolved generator synthesizes hallucination data to further optimize the detector through rule-based reinforcement learning. Experiments on RAGTruth and LLM-AggreFact across three model families demonstrate that the proposed framework can progressively enhance a small LLM to match or even outperform advanced LLMs without external supervision. Our code is available at https://github.com/maybenotime/Hallucination_Self-Play.

We report the empirical reliability of Gemini models as audio judges that score full-duplex agent conversations directly from the raw stereo waveform, tested across three models in the Gemini family: 2.5 Flash, 3.5 Flash, and 3.1 Pro. Our primary evidence base uses Gemini 2.5 Flash as the ground-truth model, validated against three calibrated human raters on 209 stereo sessions, scored on 8 production dimensions: 152 full-duplex conversations across 13 accent-and-condition strata, together with 57 adversarial defect-injected clips. The evidence for Gemini 2.5 Flash is consistent across three tests. (i) On 5 of 8 dimensions the LALM-human Spearman rho departs from the pairwise human-human rho by at most 0.07, and on 7 of 8 dimensions the two quantities 95 percent bootstrap confidence intervals overlap. (ii) The LALM agrees with the three-rater human mean within 1 point on 60 to 92 percent of sessions on 6 of 8 dimensions. (iii) On 45 of 48 (defect, dimension) cells the LALM is as sensitive as humans or better under Newcombe-Wilson 95 percent confidence intervals, though most of these are underpowered nulls rather than demonstrated parity. Rank-ordering ability transfers across the Gemini family: 3.5 Flash improves simple agreement to 8 of 8 dimensions, while 3.1 Pro rates several dimensions markedly lower than humans despite comparable rank correlation. A model swap should be re-validated on calibration specifically, not assumed from rank-correlation alone. We identify four areas where deployment requires care, and we estimate that human rating alone for our current evaluation cadence costs roughly two orders of magnitude more than the equivalent LALM workload. The data presented here provides a defensible empirical basis for deploying the LALM as a substitute or fourth rater on the dimensions where the evidence supports it.

Although reinforcement learning (RL) improves the reasoning capabilities of large language models (LLMs), GRPO-style methods can suffer from a positive-credit contamination failure mode. Specifically, uniform credit assignment can reinforce locally erroneous tail tokens alongside useful tokens within successful trajectories. To address this problem, we propose Tail-Aware Credit calibratiOn (TACO), which calibrates the positive credit assigned to such tokens without token-level error labels. To this end, TACO lets the policy serve as its own reference, assessing each sampled token's rarity relative to the policy's local expectations to compute a tail-risk score. Using this score as a proxy for potential local errors, TACO softly reduces positive credit for high tail-risk tokens while retaining nonzero weights, limiting their reinforcement without eliminating opportunities to learn useful rare patterns. Experiments across four LLMs and eight reasoning benchmarks show that TACO consistently improves over GRPO-style baselines. Notably, TACO also improves training stability, supporting sustained performance gains in long-horizon RL. The source code is available at: https://github.com/xiuyilou/TACO.

Intent detection is a critical task that bridges human intents and system actions in human-machine interaction systems. However, there still exist challenges for detecting out-of-scope (OOS) intents. (i) The traditional methods view the OOS intent detection as a multi-class classification, then the detection accuracy decreases as the class number of the known intents increases; (ii) LLM-embedding methods require large parameters, that makes them difficult to train and practically deploy. Thus, this work proposes a multi-cluster boundary learning method to detect OOS intents via MiniLM embedding (i.e., all-MiniLM-L6-v2) in an one-class classification workflow. The method learns the boundaries of multi-cluster embeddings generated by MiniLM from the training utterances, and then rejects the out-of-domain utterances as OOS intents. Experiments are conducted on public CLINC150, StackOverflow and Banking77 datasets. The results show that the method achieves the state-of-the-art OOS intent detection performance compared the other baselines. Ablation studies are also conducted and the results show that the used MiniLM can better adapt to the workflow and utterance embedding requirements. The code is available at supplementary materials.

Motion and emotion are core parts of intelligent, expressive behavior. In this paper, we introduce fog, a function composition framework for implementing and compose motion functions. We demonstrate how fog can be used to express motion and emotion in Heider-Simmel style animations. This code generation framework can help users generate functions for verbs, adverbs, gestures, and emotions to create an open-ended motion vocabulary. It is complemented by an animation editor that helps users refine motion through direct manipulation and dynamically generated UI. We evaluate our approach with a perceptual evaluation, where we test 452 fog-generated animations to see if people can recognize the semantic meaning of the motion. We find that fog's motion functions can be recognized at 68% accuracy, a 2.68x improvement over a chance baseline. In a mixed-methods user study with professionals and novices, we show that fog in interface form can support users with more rapid iteration, exploration, and control.

As Green Software Engineering matures, energy efficiency has transitioned into a mission-critical non-functional requirement. While software design patterns ensure structural integrity, their inherent abstraction layers impose an implicit "metabolic cost" that often remains obscured during the design phase. This paper empirically investigates the energy dynamics of the Memento design pattern, contrasting a direct, unabstracted baseline against Classic full-snapshot and Differential delta-encoding strategies. Leveraging the RAPL interface for high-fidelity hardware telemetry, we quantify energy dissipation across state volumes scaling from 10 MB to 200 MB. Our empirical results expose a critical architectural trade-off: the Differential strategy minimizes memory traffic, yielding a maximum energy reduction of 65.8% for mid-scale states, but collides with a catastrophic "memory wall" at 200 MB. At this saturation point, algorithmic optimizations are completely neutralized by severe GC thrashing and non-linear power spikes. We synthesize these findings into evidence-based heuristics, providing architects with a robust framework to reconcile structural design quality with sustainable Green IT imperatives.

Preprocessing-based methods for stereotype mitigation, such as pre-/post-training on debiased corpora, are widely used in NLP. While these approaches reduce measurable stereotypes for targeted groups, we find they often induce unintended shifts-side effects, where stereotyping or counter-stereotyping can increase relative to neutral baselines for other demographics, including across unrelated demographic categories. We demonstrate these side effects across two model families (encoder-only and decoder-only), multiple preprocessing strategies (removing stereotypical sentences, removing group mentions, and swapping group references), and both pre- and post-training at different data scales on Wikipedia. Standard benchmarks frequently miss these shifts. Using attention-rollout analysis, we observe that such side effects are not accompanied by large changes in attention flow, complicating mechanistic explanations. We discuss implications for evaluation, provide actionable diagnostics, and argue for side-effect-aware, transparent mitigation practices.

Current safety methods for large language models are known to be vulnerable to adversarial attacks, motivating research into robust alternatives. Latent Adversarial Training (LAT) is among the most effective defenses, but can degrade utility and requires training on large datasets of harmful prompts. We introduce Latent Personality Alignment (LPA), which replaces explicit harm refusal with adversarial training on just 66 harm-agnostic statements drawn from psychometric personality literature. We hypothesize that personality-anchored representations share latent structure with harm avoidance, so adversarially stabilizing them implicitly constrains the subspace exploited by jailbreak attacks. LPA achieves near-zero attack success rates on HarmBench across direct requests and five jailbreak methods, despite never seeing harmful content during training and no loss of performance on standard benchmarks. Moreover, the training process is lightweight; the entire procedure completes in minutes on a single GPU and uses 75x fewer examples than standard LAT. Extensive ablations demonstrate the robustness, efficiency, and generalization of our method.

Large language models (LLMs) are reshaping social science methodology. Researchers increasingly prompt language models to generate quantitative measurements of social concepts, for example labeling data or simulating survey responses. Yet LLMs pose methodological challenges including bias, hallucination, and brittleness across contexts, with unclear threats to validity. Standard practices and norms for addressing these challenges are still emerging. We collect and systematically analyze validation practices in a comprehensive corpus of papers from eight flagship social science journals that use LLMs as measurement instruments. We find that LLM-generated measurements frequently play a central role in empirical analyses, yet validation practices are inconsistent and limited. We outline complementary strategies for more robust validation, pointing toward better norms and standards around the use of LLMs in social science.

With the growing adoption of VLMs, DMs, LLMs, and AFMs, these multimodal foundation models can inadvertently encode sensitive, copyrighted, biased, or unsafe cross-modal associations that originate from their training data. Retraining after deletion requests or policy updates is often impractical, and targeted forgetting remains difficult because knowledge is distributed across shared representations. Multimodal unlearning addresses this challenge by enabling selective removal across modalities while retaining overall utility. This survey offers a unified, system-oriented view of multimodal unlearning across vision, language, audio, and video, grounded in recent advances, emerging applications, and open problems. Our taxonomy enables systematic comparison across model architectures and modalities, clarifying trade-offs among deletion strength, retention, efficiency, reversibility, and robustness. This survey highlights open problems and practical considerations to support future research and deployment of multimodal unlearning. We release a curated repository: https://smsnobin77.github.io/Awesome-Multimodal-Unlearning/

Research on stereotypes in large language models (LLMs) has largely focused on English-speaking contexts, due to the lack of datasets in other languages and the high cost of manual annotation in underrepresented cultures. To address this gap, we introduce a cost-efficient human-LLM collaborative annotation framework and apply it to construct EspanStereo, a Spanish-language stereotype dataset spanning multiple Spanish-speaking countries across Europe and Latin America. EspanStereo captures both well-documented stereotypes from prior literature and culturally specific biases absent from English-centric resources. Using LLMs to generate candidate stereotypes and in-culture annotators to validate them, we demonstrate the framework's effectiveness in identifying nuanced, region-specific biases. Our evaluation of Spanish-supporting LLMs using EspanStereo reveals significant variation in stereotypical behavior across countries, highlighting the need for more culturally grounded assessments. Beyond Spanish, our framework is adaptable to other languages and regions, offering a scalable path toward multilingual stereotype benchmarks. This work broadens the scope of stereotype analysis in LLMs and lays the groundwork for comprehensive cross-cultural bias evaluation.

Roy Harris's Integrationist linguistics offers a compelling critique of the referentialist tradition embedded deep at the heart of computational approaches to language, arguing that language is not a code that maps onto a pre-given world but a situated, bipartite activity oriented toward prospective joint action. Yet Integrationism leaves certain explanatory gaps: it does not fully account for the structural mechanism by which signs sustain prospective openness, it undertheorises the continuity between linguistic and non-linguistic semiotic activity, and it offers no detailed account of the structural properties of the accumulated archive of past integrations. This paper argues that Elan Barenholtz's autogenerative theory of language, developed in response to the behaviour of Large Language Models (LLMs), can fill precisely these gaps, enriching Integrationism without undermining any of its core commitments. Specifically, the autogenerative account provides: a structural mechanism for the prospective openness that Harris identifies as central to bipartite communication; a computational correlate for Harris's thesis of semiotic continuity between language and other sign-making activity; and a theory of the archive: what the accumulated residue of past integrations looks like and how new participants draw upon it. The synthesis preserves Harris's ontological primacy of the situated integrative act while adding explanatory content that Integrationism itself does not supply. For practitioners and researchers in natural language processing and large language model design, the argument offers a principled account of what the statistical structure that LLMs so effectively exploit actually is, and of what it cannot, by its nature, provide.

Training tool-use agents to improve from their own experience remains challenging, as supervised fine-tuning relies on fixed teacher-distilled trajectories, while sparse-reward reinforcement learning provides weak supervision for long-horizon interactions. We present DeepSearch-Evolve, a self-distillation framework for web agents built on DeepSearch-World, a deterministic and verifiable environment with reproducible search and page-reading tools. DeepSearch-World contains 420K multi-hop QA tasks constructed from entity-level random walks and supports key agentic cognitive behaviors useful for self-evolving, including progress verification, grounded reflection, and failure recovery. DeepSearch-Evolve iteratively performs trajectory generation, filtering, data mixing, and fine-tuning to train stronger agents. Without distillation from more capable models, DeepSearch-World-9B achieves competitive performance compared with open-source agents, reaching 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, showing that verifiable environments enable scalable self-evolution for long-horizon web agents. We will release the environment, 420K training pool, validation set, model, and code to facilitate future research on self-improving deep search agents.

Recent developments in AI for Mathematics (AI4Math), especially Large Language Model (LLM)-driven theorem provers, has achieved remarkable success in formal proof generation for well-defined mathematical problems through Interactive Theorem Proving (ITP) languages. However, current systems remain fundamentally limited in tackling frontier research mathematics, such as discovering new theorems or resolving open conjectures, which are often open-ended, under-specified, and involve multiple layers of abstraction. We argue that the next leap in AI4Math systems requires a decisive shift from predefined problem-solvers to research agents that can address frontier mathematical challenges with rigorous formal mathematical reasoning. In this position paper, we provide a systematic review of the field, covering datasets, auto-formalization, and proof synthesis. More importantly, we identify core limitations of existing systems in serving as mathematical research agents, examining issues across datasets, relational structure, mathematical exploration, tool ecosystem, and human-AI collaboration, outlining a strategic road-map for the future of AI4Math.

In this age of social media, sites like Twitter have become meeting places for people to share their views and feelings on a wide range of issues and current events as they unfold in real time. Sentiment analysis, a critical application of NLP, has become indispensable due to the massive influx of user-generated content, enabling the extraction of meaningful insights from the opinions and emotions expressed in textual data. Sentiment analysis on Twitter employs sophisticated computational techniques to categorize tweets into positive, negative, or neutral sentiments. This method not only examines individual expressions but also analyzes vast databases related to specific subjects or events. By spotting these emotions, machine learning models help improve public opinion interpretation and trend forecasting. This paper examines the effectiveness of various machine learning and deep learning approaches. Designed for this use, the system evaluates logistic regression, random forest, naïve bayes, gradient boosting, and LSTM networks, among other algorithms applied in sentiment classification. This work identifies the optimal sentiment analysis model using a Kaggle Twitter dataset that has been preprocessed through tokenization, lemmatization, and stopword elimination. Emphasizing the better performance of the LSTM approach, the model attained a training accuracy of 90.98%, a testing accuracy of 80.00%, and a micro-average ROC- AUC score of 0.92. These results show that the model outperforms conventional machine learning techniques in capturing contextual and sequential textual aspects.

Structure-property relationships are foundational to biology, chemistry and materials science, where function, reactivity and physical response emerge from spatial, chemical and periodic organization. Mechanistically explaining these relationships requires interpreting structural evidence through scientific principles and physical constraints, from stereochemistry and bonding to symmetry, energetics and periodic order. However, applying artificial intelligence to this process presents a joint challenge of representation and reasoning: models must preserve domain-native structural information while showing how specific evidence supports predictions under these constraints. Here we introduce SciReasoner, a multimodal scientific foundation model for native structural reasoning across proteins, small molecules and inorganic crystals. SciReasoner discretizes coordinates, topologies and periodic connectivities into a unified structure-aware vocabulary, treating structural tokens as addressable evidence units during reasoning. In homology-controlled Gene Ontology prediction, SciReasoner improves Cellular Component annotation for low-homology and orphan-like proteins, increasing $F_{\max}$ from 0.42 to 0.55. In chemistry, it raises single-step retrosynthesis accuracy from 0.63 to 0.72 while generating fragment-level disconnection and precursor-verification traces. In materials science, its representations separate elemental and compound phases and resolve high- and low-band-gap regimes. Across 86 benchmarks, SciReasoner achieves state-of-the-art performance on 67 tasks. Double-blind expert evaluation rates its reasoning traces as preferred or at least comparable to those of a frontier large language model in 98% of cases. By making structure an inspectable substrate for reasoning under scientific constraints, SciReasoner connects accurate prediction with interpretable scientific inference.

Externalizing knowledge in LLM pre-training is a promising avenue to achieve higher performance at smaller scales, control knowledge use, and overall increase model transparency. We propose continuous-query limited memory language models (Co-LMLM), an LLM that interleaves flexible vector retrieval queries with next-token predictions, and is pre-trained to copy knowledge returned from the KB, rather than memorize it. Co-LMLM is pre-trained with a scalable approach that jointly trains a knowledge-externalizing LLM, induces its knowledge base, and learns an expressive continuous retrieval mechanism. Across pre-training at multiple model scales, Co-LMLM outperforms prior knowledge-externalizing and vanilla LLMs in both perplexity and factual precision. At 360M scale, this includes lower perplexity than models pre-trained on 40$\times$ more data, and SimpleQA-verified performance that is in line with gpt-4o-mini and higher than Claude Sonnet 4.5.

The optimization of long-horizon agents increasingly relies on reflection-based mechanisms, where a large language model (LLM) acts as an optimizer to diagnose agent failures and improve agent policies. However, real execution traces are difficult to use directly for optimization: large trace collections are often redundant and heterogeneous, making optimization inefficient and prone to overfitting to low-value failures; meanwhile, each individual trajectory also contains many irrelevant steps, while naive context reduction methods such as truncation or sliding windows can discard causally important evidence and produce misleading optimization signals. To resolve this dilemma, we introduce STRACE (Structural TRajectory Analysis and Causal Extraction), a framework that constructs high signal-noise optimization contexts for more precise and effective optimization. At the batch level, STRACE mines failure patterns to filter redundant traces and retain representative failures; within each selected trace, it performs causal localization over a textual dependency graph to remove non-causal steps and identify the true root-cause module for optimization. Empirical results demonstrate that STRACE significantly outperforms standard context-filtering baselines. Notably, on a challenging formal verification task (VeruSAGE-Bench), it successfully optimizes human-expert designed agents, delivering $1.4\times$ success-rate improvement (42.5% to 58.5%). The code is available at https://github.com/moomight/STRACE .

Reinforcement learning from verifiable rewards (e.g. GRPO) is the engine behind today's reasoning models, yet it grades only the final answer. On hard problems this trains models to write more rather than to think better, since the trace itself is never graded and no label for good thinking exists. We introduce Agon, which makes two competing models each other's graders. Both attempt the same problem; in alternating roles, one drafts a solution and the other reads it while solving, and each is rewarded for out-solving the other. To win, a model must out-reason a rival that has seen its work, so reasoning is judged implicitly during training, with no process labels and no reward model. Because both models are optimized, each faces a progressively stronger rival, which single-model RL cannot provide. The two need only be comparably strong and behaviorally different. At inference the pair deploys as it trains, a two-stage cascade in which one model drafts and the other answers after reading the draft. On the hard split of DeepMath with Qwen3, this doubles GRPO's pass@1, roughly eight times the gain of an untrained Mixture-of-Agents pass over the same base. The ordering replicates on competitive-programming code and across model families (Qwen3.5, Gemma 4). For now the models talk in text; the next step is to let them reason together in latent space.

Group Relative Policy Optimization (GRPO) stalls on a model's hardest problems: when no rollout in a group succeeds, the group-relative advantages vanish and the problem contributes no gradient, wasting the frontier examples we most want to learn from. Prepending a correct prefix of a reference solution raises the success rate, making prefix length a continuous knob on difficulty. Concurrent methods set the knob once; AdaPrefix-GRPO turns it into a feedback controller: throughout training it adjusts how much of the solution each problem gets, holding its success rate near 50%, where GRPO's gradient signal is largest, then withdraws the assistance entirely, so the deployed model solves problems unaided. On hard math, at matched training FLOPs, it more than doubles GRPO's accuracy on held-out problems from the training distribution for a 0.6B model (2.1x), with 1.6x on Qwen3-1.7B and 1.7x on AIME, while roughly halving trace length. The method is implemented in data preparation plus a loss mask on prefix tokens; the trainer is otherwise stock. The smaller the model, the larger the gain.

Large language models hallucinate most about entities they have never seen. We ask whether a model's activations betray entity familiarity before a single answer token is generated, and whether that signal predicts the factual reliability of the answers. On four Polish Bielik models (1.5B-11B parameters), we probe four entity domains (athletes, cities, writers, musicians), each with 42 well-known, 42 obscure-but-real, and 42 fabricated entities addressed by a one-sentence question (504 prompts per model). Two unsupervised, single-forward-pass dispersion measures over post-SwiGLU MLP activations, inverse participation ratio and spectral entropy, separate known from fabricated entities at AUROC 0.95-1.00 across all domains and scales; a supervised linear probe reaches 0.99-1.00. Both clear selection-aware permutation floors of about 0.70-0.74 (empirical p<=1e-3), survive held-out layer selection (0.93-0.99), and persist on real names (known vs. obscure-but-real: 0.96-1.00). The signal transfers across entity types (mean off-diagonal AUROC 0.92-0.99); a matched-template counterfactual shows the only large drops are template-caused, not entity-type effects, and the signal is diffuse across heads. This representational signal is already at ceiling at 1.5B, whereas behavioral factual reliability scales sharply: 0, 2, 10, and 19 of 42 known athletes are answered fully correctly by the 1.5B, 4.5B, 7B, and 11B models under a strict judge. Within known entities, separating correct from hallucinated answers is much harder (probe 0.93; dispersion no better than a first-token-entropy baseline). A five-sample semantic-entropy baseline reaches only 0.71-0.83 at 5x the inference cost. Despite this internal awareness, the models almost never abstain: an audit of 2,520 answers finds 2 refusals and 1 hedge. Entity familiarity and factual reliability are distinct phenomena on different scaling curves.

Large language models increasingly understand dialectal English, yet still produce only standard, US-leaning English, leaving dialectal generation, the harder half of the problem, largely unaddressed. We introduce DiaLLM, which continually pretrains three open-weight language model families on the International Corpus of English and applies implicit and explicit post-training paradigms, each combined with three model alignment strategies, giving the first controlled comparison of these components across Australian, Indian, and Northern British English. Our results reveal a robustness-generation gap: benchmarks are shaped by continual pretraining and SFT, while alignment visibly reshapes generation in ways benchmarks do not capture. Explicit variety-targeted adaptation produces output reliably recognised as dialectal and judged more dialectal than broad alignment, yet where human judgement was directly assessed, the method that most aggressively optimises the dialectal reward is not the one judged most dialectal. Independent linguistic analysis corroborates this reward-quality gap, most clearly on two of the three families. No single alignment method dominates, and closing the gap will require richer reward designs and continued investment in dialectal resources. We release all code, checkpoints, and preference datasets.

Does RL post-training merely amplify primitive skills already latent in a base model, or can it compose primitive skills into new higher-level strategies? We study this question in a fully observable rewrite-grammar environment where the pretraining distribution is known and every generated rewrite can be audited. A Transformer is pretrained on primitive symbol-rewrite chains and post-trained on a Trace-based reasoning task with only a binary final-answer reward. RL solves held-out problems that remain rarely solved by the pretrained model even under much larger sampling budgets, while rejection fine-tuning improves early but plateaus. Trace analysis shows that RL reorganizes primitive competence through a phased compositional mechanism: it first strengthens primitive reductions, then discovers valid composed procedures. These include sequential compositions, which collapse ordered chains of primitive contractions, and parallel compositions, which combine independent primitive contractions in a single step. The composed procedures are not isolated samples; they are reused and consolidated into a stable repertoire. Comparing RL with rejection fine-tuning shows that the key difference is not exploration volume but selectivity: RFT produces many shortcut-like rewrites, much of them invalid, whereas RL concentrates exploration into valid reusable structure. Pretraining ablations show that the emergence of compositional strategies is gated not by primitive exposure alone, but by whether pretraining organizes primitive competence into reduction procedures that RL can later compress. The base model provides weak procedural ingredients; RL builds them into reliable higher-level strategies.