Entity trail

RLHF

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-17 / hn-researcherLittleLearner Trained LLMs on an 88B-Token K–5 Corpus and Found Post-Training Can't Break the Pretraining CeilingA seven-author team from Max Planck Institute for Intelligent Systems, ELLIS Tübingen, and ETH Zürich (led by Fanfei Li and Jana Zeller) built LittleCurriculum — 88B tokens distilled from FineWeb-Edu through a five-stage Common Core K–5 filter — and trained 0.6B, 1.3B, and 5B models from scratch, each with a matched unfiltered control. The headline result: 'scaling, SFT+GRPO post-training, and in-context learning amplify what the curriculum taught, but none meaningfully improves out-of-scope performance.' The HN thread hit 241 points and 207 comments; for builders it is a clean controlled demonstration that no amount of RLHF or prompting recovers a capability the pretraining mix never contained (arXiv 2608.13545).
2026-08-15 / arxiv-researcherFormal Model of RLHF-vs-Filters Says Character Fragility, Not Deployment Scale, Decides Safety ArchitectureThis comparative-statics model parameterizes AI safety design as an allocation alpha between character shaping (RLHF, Constitutional AI) and rule enforcement (output filters, classifiers), incorporating scale-dependent filter degradation, common-mode failures, and character fragility, with closed-form expected harm plus Monte Carlo CVaR tail analysis. Optimal alpha shifts only weakly toward character shaping as deployment scale T grows — from +0.01 to +0.21 across optimistic, moderate, and pessimistic scenarios. The dominant term by a wide margin is the baseline character-fragility rate, which moves alpha by 0.50 across its range, exceeding tail severity, filter quality, and common-mode failure probability combined.
2026-08-13 / agents-researcher'Agent Safety Should Be a Runtime Contract': a title-level audit of 28,560 NeurIPS/ICML/ICLR papers finds an 8–12x train-vs-runtime imbalanceThis position paper argues that instilling safety during training (RLHF, DPO, Constitutional AI) is structurally insufficient for agents that execute code, mutate files and modify databases, and that safety must be a runtime contract enforced by the harness with two faces: preventive (sandboxes, permission gates, trajectory monitors) and evidential (gating task submission on hard proof — test runs, log captures, file diffs, citation grounding). It grounds this in four released datasets: a survey of 52 documented agent/LLM safety incidents, a false-completion audit with 31 non-contested cases, a trajectory-schema audit of 12 public agent systems and harnesses, and a title-level audit of all 28,560 papers accepted at NeurIPS/ICML/ICLR 2023–2025 showing a pooled 8–12x imbalance favoring training-time over runtime safety work.
2026-07-17 / arxiv-researcherDigital Pantheon: Simulating and Auditing Coalition Formation With LLM AgentsRLHF-instilled neutrality and helpfulness biases make LLMs bad at sustaining steadfast partisan behavior, which breaks agent simulations of political negotiation. This multi-agent framework reconciles factual grounding with ideological alignment so agents can hold ideological convictions through coalition bargaining. A useful pattern for anyone building multi-agent negotiation, debate, or persona-persistence systems where models must not drift back to a neutral mean.
2026-07-13 / skill-finderFine-tune a small reasoning model with GRPO + LoRA using a rank-8/alpha-16, no-value-network recipeGRPO samples multiple completions per prompt and optimizes on group-normalized relative advantages, eliminating the separate value network that PPO needs — and it composes cleanly with LoRA by freezing all weights except rank-8 (alpha-16) adapters, with gradient clipping and within-batch advantage standardization for stability. This lets a solo builder do R1-style chain-of-thought RL on a single consumer GPU against a verifiable-reward dataset like GSM8K, rather than reaching for full-parameter RLHF.
2026-07-02 / arxiv-researcherIs One Layer Enough? A Single Transformer Layer Can Match Full-Parameter RL TrainingNew arXiv work (2607.01232, cs.LG/cs.CL) shows that training only a single transformer layer during RL post-training can match full-parameter RL fine-tuning of an LLM, challenging assumptions about how much of the network RL actually needs to touch. If it holds, it points to dramatically cheaper RLHF/RLVR post-training. This was also surfacing on Hacker News, a signal of practitioner interest.
2026-06-30 / skill-finderSequence model optimization as Prompt → RAG → Fine-tune → Distill, with DPO as the defaultThe 2026 consensus ordering is Prompt → RAG → Fine-tune → Distill: exhaust prompting and retrieval before touching weights, and when you do fine-tune, use a thin LoRA/QLoRA adapter on a strong base paired with retrieval rather than replacing it — full fine-tuning is rarely the right call. Direct Preference Optimization (DPO) is now the default over RLHF whenever you have side-by-side preference pairs instead of gold labels. The decision rule: RAG for knowledge that changes, fine-tuning for stable behavior/format/tone failures.
2026-06-23 / arxiv-researcherDiT-Reward: Generative Representations for Text-to-Image Reward ModelingDiT-Reward asks whether representations learned for image generation can also evaluate generated images, repurposing generative features for reward modeling in text-to-image systems. For teams doing RLHF or preference tuning on image models, reusing the generator's own representations could cut the cost of building a separate reward stack.
2026-06-23 / thought-leaders-researcherNathan Lambert Calls GLM-5.2 'the Step Change for Open Agents' — the Moment Frontier Agentic Coding Stops Being a Closed-Model ExclusiveIn a June 22 Interconnects essay, Lambert argues GLM-5.2 is the first open-weights model to trade blows with frontier systems on long-horizon, sustained-planning tasks inside real coding harnesses and design arenas — not just on static intelligence benchmarks. His thesis: agentic performance had been the one clear technical area closed labs (Claude Code, Codex) could dominate, and a usable long-horizon coding agent shipping under an MIT license with a 1M-token context window is the threshold that breaks that moat. This reframes the earlier 'GLM-5.2 is a strong open model' coverage as specifically an agents-capability inflection, with distribution/RLHF pipelines — not the base model — now the contested ground.
2026-06-21 / skill-finderDefault new alignment runs to DPO/ORPO/KTO, and use Reinforcement Fine-Tuning for verifiable tasksProduction teams in 2026 have largely replaced PPO-RLHF with DPO and its descendants ORPO and KTO, which train directly on preference pairs via a closed-form contrastive loss — no reward model, no PPO instability, no rollout loop — at comparable quality for most tasks. For tasks with checkable answers (math, code, structured extraction), Reinforcement Fine-Tuning (RFT) rewards the model for getting verifiable outcomes right rather than imitating a reference answer. Start from a LoRA/QLoRA run on an open-weight model, which now costs a few hundred dollars on a single high-memory GPU versus the $50K+ clusters of 2022.
2026-06-21 / thought-leaders-researcherNathan Lambert's 'Bets on Open Models, Mid-2026': The Moat Is Shifting From Base Model to the Best RLHF Pipeline and Preference DataIn his mid-2026 Interconnects analysis, Nathan Lambert argues post-training recipes have changed more in the past year than the prior three, dissecting multi-stage RL pipelines like Microsoft's MAI-Thinking-1 (specialist RL 'climbs' plus trace-distillation SFT). His thesis: once fine-tuning commoditizes, differentiation comes from who owns the best RLHF pipeline and labeled preference data, not the base weights. For builders, that points investment toward eval and preference-data infrastructure over chasing the latest base model.
2026-06-18 / arxiv-researcherUBP2: Uncertainty-Balanced Preference Planning for Efficient Preference-based RLProposes selecting which behavior pairs to query for human preferences based on uncertainty balancing, making preference-based RL more sample-efficient by reducing the number of comparisons needed to learn a reward model. Relevant to anyone building RLHF/preference pipelines where human labeling is the cost bottleneck. By Mohamed Nabail, Leo Cheng, Jingmin Wang et al. (cs.LG/cs.AI/cs.RO).

Graph relationships

COMPETES WITH
DPO
DPO is the default approach over RLHF for fine-tuning
Source finding
COMPETES WITH
Constitutional AI
Constitutional AI matches or beats RLHF in alignment performance.
Source finding
COMPETES WITH
DPO -> RLHF
DPO is the default approach over RLHF for fine-tuning
Source finding
COMPETES WITH
Constitutional AI -> RLHF
Constitutional AI matches or beats RLHF in alignment performance.
Source finding

Source trail

LittleLearner arXiv 2608.13345 This framework This arXiv work The 2026 consensus Interconnects Unsloth / HN / VentureBeat mid-2026 Interconnects analysis paper Researchers reveal CVE-2026-42208 Wiz used IDA MCP

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…