Looking for Ricursive (the AI chip design company)? You want ricursive.com|Looking for Recursive AI / Recursive Superintelligence (Richard Socher's startup)? You want recursive.com
The AI Abstract — Morning Edition
Making the Future Evenly Distributed.
A 3.8B parameter model now outperforms larger AI systems at improving other AI systems, using a trick that cuts the cost of that feedback loop by a third.
A language model is now being trained to make other language models better. Not in a vague, aspirational sense. Researchers have built a working system where a 3.8B parameter model learns, through trial and error, to generate improvement proposals for other AI models, and it beats larger models at this task while cutting the cost of generating training feedback by 33%.
The paper is 🔬LM2LM, and the mechanism is worth understanding because the obstacle it solves is the one that has made recursive AI self-improvement mostly theoretical until now.
Here's the problem: to train a model to improve other models, you need feedback on whether its improvement suggestions actually worked. The obvious way to get that feedback is to run the suggested improvements through real training and measure what happened. That's prohibitively expensive at scale. The cheaper alternative is to use a surrogate model to predict whether an improvement would work, but surrogate models tend to break down when the suggestions get far enough from the training distribution. Think of it like using a wine expert to judge wines from a completely different region than they trained on. The scores stop being reliable.
LM2LM solves this with what the researchers call a "frozen research world model," a separate model held fixed during training whose job is to simulate what would happen if an improvement were applied. Because it's frozen, it stays stable as a reward signal. Because it's separate from the model being improved, it doesn't collapse under distribution shift the way a jointly-trained surrogate would. The reinforcement learning loop then trains the 3.8B proposer to generate suggestions that this world model rates highly, without ever needing to run expensive real training runs to check.
The result is a smaller model outperforming larger baselines on improvement proposal quality, with a 33% reduction in data generation cost. Those numbers matter independently. The size reversal matters because it suggests the task of "figure out how to make a model better" is learnable in a focused way that doesn't require general capability at scale. The cost reduction matters because it's what makes the loop tractable enough to run repeatedly.
The broader implication is the one worth sitting with: this is a concrete, working step toward AI systems that automate AI research. The "models" signal in our tracker has now accumulated 236 mentions since May, and the through-line is the same each time: the field is building the infrastructure for models to participate in their own development. LM2LM is the clearest empirical demonstration of that yet.
A separate result this edition challenges a widespread assumption in two different domains at once. Two papers, clustered on the "scaling" signal, both find that bigger models aren't actually necessary for tasks where everyone assumed they were.
🔬BEST, a new framework for brain-computer interface speech decoding, matches state-of-the-art performance on Brain-to-Text benchmarks using 96% fewer parameters than comparable methods. Brain-computer interfaces translate neural signals into text, and the assumption has been that decoding the noisy, high-dimensional signals coming off an electrode array requires heavy model capacity. BEST shows the binding constraint isn't model size but cross-modal alignment: how you bridge the gap between brain signal space and language space. Once that alignment is handled well architecturally, a small model does the job.
In audio-language models, 🔬ConvexLLM takes a different angle on the same underlying problem. Adding audio understanding to a language model usually degrades the text capabilities the model already had, because the new acoustic representations interfere with the existing ones. ConvexLLM's C-Gate interface addresses this by forcing audio representations to be expressed as weighted combinations of the model's existing text embeddings, as if you had to describe every sound using only words the model already knew. The model can't learn audio representations that break its text knowledge, because audio is only allowed to speak the language of text. It hits 75.36 on VoiceBench aggregate, best among open-weight audio-language models, with the strongest gains on reasoning and knowledge tasks where text capability matters most.
If you're using a language model as a judge to evaluate other models, a new paper has a specific problem you need to know about. 🔬This study ran a controlled experiment where the thing being evaluated wasn't generated text at all, just whether a narrative followed a set of constraints. The only variable was whether the model was told it had written the content, or that another model had. That label alone shifted scores, in both directions, even when blind evaluation removed the usual self-preference effect.
The mechanism is distinct from the well-documented "models prefer their own outputs" bias. This is about authorship attribution as a concept: once a model is told something was written by a particular author, that framing changes the evaluation, regardless of which direction the label points. The practical consequence is that any benchmark using LLM-as-a-judge for comparative model ranking has a confound that isn't controlled for by blinding alone. The ranking your evaluation system produces may be as much a function of how you labeled the outputs as what the outputs actually contained.
Two papers address the infrastructure of tool-using AI agents, and they're worth reading together. 🔬Actions with Receipts introduces a cryptographic framework that binds an agent's claims, its cited sources, its execution logs, and the software version it was running into a single tamper-evident package. The attack it defends against is subtle: an adversary who can take a valid citation and a valid execution log from one context and transplant them onto a different claim in a different context. The framework detected 99.61% of these cross-object substitution attacks across 1,280 test cases.
🔬ContractRL addresses a lower-level but chronic failure mode: malformed tool calls. When an agent generates a JSON instruction to call an external tool and gets the structure wrong, current systems usually regenerate the entire object from scratch. ContractRL trains a model to repair only the broken parts incrementally, like editing a sentence rather than rewriting the paragraph. The result is 93.6% semantic success with 34% fewer tokens than full regeneration. Both papers reflect the same underlying pressure: as agents get deployed in higher-stakes settings, the field needs the kind of auditability and reliability guarantees that software engineering has had for decades.
Two more results round out the picture. 🔬MuninAI is an open-source, self-hosted framework that raises LLM accuracy on specialized domain literature from 38.7% to 87.4% by wrapping a model in purpose-built reading and search tools rather than swapping in a bigger model. The performance gain comes entirely from tool architecture, not base model capability, which matters for any research group that has proprietary corpora they can't send to a commercial API.
🔬NLData2Opt is a new benchmark that tests whether LLMs can formulate optimization problems from natural language descriptions and messy real-world data, extracting parameters, computing derived values, and filling in gaps. Models that score well on cleaner existing benchmarks drop substantially when evaluated here. The gap quantifies something practitioners already suspect: lab evaluations are running on cleaner inputs than production ever will.
One note on coverage: today's payload is entirely preprints. These results are not yet peer-reviewed in the traditional sense. The empirical claims are specific enough to be worth taking seriously, but none have cleared independent replication.
🔬 LM2LM: Read for the world model mechanism and what it means that a 3.8B model can outperform larger ones on improvement proposals.
🔬 Self- and Other-Labels Induce Bidirectional Bias in LLM Judges: Read for the experimental design, which isolates authorship attribution as a confound in ways that have immediate implications for anyone running model comparisons.
🔬 Actions with Receipts: Read for the threat model, specifically the cross-object substitution attack, which is the kind of thing that will matter more as agents move into regulated domains.
🔬 MuninAI: Read for the tool architecture breakdown and what it tells you about where performance actually comes from in domain-specific deployments.
🔬 BEST: Read for the cross-modal alignment argument, which applies well beyond brain-computer interfaces to any task where you're bridging two very different signal types.
Links
- Language Models Learn to Improve Language Models
arxiv.org
Researchers propose LM2LM, a system that trains LLMs via reinforcement learning to generate effective improvements to other language models, using a frozen 'research world model' as a scalable reward signal instead of expensive real training runs. The work addresses both the computational cost of feedback collection and the reliability problem of distribution shift in surrogate rewards, showing a 3.8B model outperforming larger baselines on improvement proposals while reducing data generation costs by 33%—a concrete advance toward automating AI research itself.
- MuninAI: A Self-Hosted Agentic Framework for Independent Research Groups
arxiv.org
MuninAI is an open-source self-hosted framework that improves LLM accuracy on domain literature from 38.7% to 87.4% by wrapping models in an agentic harness with specialized reading and search tools, rather than relying on naive retrieval. The system keeps proprietary research corpora local while demonstrating that the tool architecture—not the base model—drives performance on specialized tasks, with evaluation suite included for reproducible adoption.
- NLData2Opt: Benchmarking Parameter Derivation and Prediction for Modeling Optimization Problems from Natural Language and Data
arxiv.org
Researchers introduced NLData2Opt, a benchmark that evaluates LLMs on multi-stage reasoning required to formulate optimization problems from natural language and noisy data—extracting relevant parameters, computing derived values, and predicting missing data. The work reveals substantial performance drops compared to idealized benchmarks, signaling that existing evaluation methods overestimate LLM capability for practical problem formulation tasks.
- Learning User Representations to Improve Simulation Fidelity
arxiv.org
Researchers propose hindsight distillation, a training procedure that generates improved natural-language user representations to condition simulators for more faithful reproduction of diverse user behaviors. The approach addresses a critical gap in AI assistant evaluation—current user simulators generalize poorly to underrepresented users—and demonstrates measurable fidelity improvements across 700 conversations, with implications for scaling interactive training of AI systems.
- ConvexLLM: Scaling Convex-Hull Speech Interfaces to General-Purpose Audio Intelligence
arxiv.org
ConvexLLM introduces a C-Gate interface that grounds audio representations as convex combinations of frozen text embeddings, achieving 75.36 on VoiceBench aggregate—highest among open-weight audio-language models. This work addresses a core challenge in multimodal model design: how to preserve pretrained text capabilities while effectively incorporating acoustic information, with strongest gains on reasoning and knowledge tasks.
- Do Speech BCIs Need Larger Models? Rethinking Neural Decoding beyond Scaling
arxiv.org
A new arXiv paper (BEST framework) challenges the assumption that large models are necessary for speech BCIs, achieving state-of-the-art results on Brain-to-Text benchmarks while using 96.15% fewer parameters than comparable methods. For practitioners working on constrained neural-interface systems, this offers both practical deployment gains and theoretical insights into cross-modal alignment and decoder scaling dynamics in low-resource settings.
- Self- and Other-Labels Induce Bidirectional Bias in LLM Judges
arxiv.org
Researchers used a controlled task (evaluating narrative constraints rather than generated text) to isolate authorship attribution as a distinct driver of LLM judge bias, finding that self/other-labels alone shift scores bidirectionally even when blind evaluation removes traditional self-preference. This matters because LLM-as-a-judge is becoming standard infrastructure for model evaluation and ranking; this work quantifies a confound that can distort comparative benchmarking and highlights methodological safeguards needed for reliable evaluation systems.
- Actions with Receipts: Jointly Binding Claims, Evidence, and Execution for Replayable Tool-Agent Auditing
arxiv.org
Researchers introduce a cryptographic "claim-anchored execution contract" that jointly binds agent claims, source citations, execution traces, and version state—preventing substitution attacks where valid citations and logs are transplanted across different claims or executions. The framework detects 99.61% of cross-object attacks and separates integrity verification from semantic validation, establishing a new standard for auditability in tool-using agents critical as these systems move into high-stakes deployment.
- Cultural (In)Competence: Investigating Culture-Specific Stereotype Detection Capabilities of LLMs
arxiv.org
Researchers systematically evaluated how well LLMs detect culture-specific stereotypes across US, Dutch, and Chinese contexts, finding that models overpredict stereotypes and fail to reliably account for cultural nuance—with non-English prompts and instruction tuning partially mitigating the problem. This work signals a concrete fairness vulnerability in cross-cultural deployment of LLMs and provides empirical data for practitioners and safety researchers optimizing model behavior across linguistic and cultural boundaries.
- ContractRL: Shielded Group-Relative Policy Optimization for Auditable Tool-Call Repair
arxiv.org
ContractRL introduces a contract-constrained reinforcement learning protocol for repairing malformed structured tool calls (JSON) incrementally rather than regenerating full objects, achieving 93.6% semantic success with 34% fewer tokens than full regeneration and statistically significant gains over supervised baselines. This matters because tool-use reliability and auditability are critical bottlenecks for LLM-based agentic systems in production environments, and the approach demonstrates how RL with structured constraints can improve both efficiency and verifiability of model behavior.