Post-training: accepted in review but not delivered (85)¶
These entries passed an earlier review and lack a current catalog entry or an explicit cost closure. This is not user authorization to implement every item, nor does it mean each item passed the current cost gate. Do not use this as an automatic implementation queue. Sorted by paper ID descending.
Confirmed decisions not to reproduce full experiments are marked separately. Other items first need cost and scope reduction, not automatic execution.
| ID / paper | Current disposition | Priority organization verification | Prior review rationale |
|---|---|---|---|
| 2610.20413 When Does Momentum Help in Reinforcement Learning with Verifiable Rewards? | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accept only the original ten-arm bandit's bias-corrected EMA, safe η/(1−β) mechanism, and analytic property checks, preserving the objective in CPU numerics. Excludes 4B/32B RLVR and the full 50,000-replicate statistics campaign; no LLM improvement claim. Accepted as a low-cost candidate, not implemented or run here and not automatically queued for execution or evolve. |
| 2610.20303 SpectrumQuant: Spectrum Alignment for LLM Quantization | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | SpectrumQuant builds closed-form alignment and mixing rotations from calibration activation covariance. The paper reports about 0.26 GPU hours and 4.51 GiB for 7B calibration; a faithful 7B path can retain 128 real calibration samples. No 70B or full downstream model sweep promised. Accepted as a low-cost candidate, not implemented or run here and not automatically queued for execution or evolve. |
| 2610.08778 Sherpa: Teaching LLMs to Teach Adaptively | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. A turn-wise gate masks returns with leave-one-out normalization among trajectories reaching the same turn. The teacher learns student preferences. MATH training and evaluation are separated, with a teaching benchmark. |
| 2610.08773 AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Freeze the world model and judge; train curriculum–executor then adversary–executor. Three rollout seeds and 150 tasks are not production A/B. |
| 2610.08670 Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Public evaluation: judgment–action gap screening, four-level calibration, and pressure controls independent of training. Apache-2.0 DeepSteer is accessible. Only 74/208 main scenarios passed screening; conditional results do not generalize automatically. |
| 2610.08108 Enhancing Diffusion Language Models with Autoregressive Post-Training Weights | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. A2D task-vector transfer and AR/diffusion delta interpolation have checkpoint benchmarks. Coefficients used the anchor benchmark; local reproduction must isolate validation selection. |
| 2610.07898 FC-SWE: Failure-Conditioned RL for Long-Horizon Software Engineering Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Real R2E-Gym training resets after failure and supplies patch/verifier context. Attempts receive separate rewards; unexecuted padding contributes no advantage. SWE-bench is verifier-assisted, not standalone pass@K. |
| 2610.07767 TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Choose adjacent training activation FP4 codewords closer to rollout values and cache mantissa/scale. Tested with four MoE trainings. Native NVFP4 needs compatible hardware; an A100 simulation cannot claim native speed. |
| 2610.07739 Cite What You Explore: Budget-Aware LLM Reasoning over Medical KGs with Verifiable Evidence | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Snapshot-matched loss differences with and without evidence define paired gain, with budget and citation integrity. Train the retrieval reasoner with REINFORCE. Needs real MIMIC/PrimeKG mappings, not hand-written scores. |
| 2610.07731 Learning to Retrieve via Reinforcement Learning in Embedding Space | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. vMF embedding policy, G×G product rollouts, RLOO, and CMP projection denoising. ReasonRank training, BRIGHT evaluation, and three paired seeds. |
| 2610.07654 Does On-Policy Distillation for Safety Pose Backdoor Risks? | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Lazy Defense clips sampled teacher–student token log-ratios on both sides and detaches them. Four combinations only delayed backdoors; they do not guarantee removal. |
| 2610.07349 RELACE: retrospective likelihood-based action credit estimation for long-horizon language agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Correct local returns with ratios between original and outcome-conditioned action likelihoods. Terminal outcomes are training credit only. Real ALFWorld/WebShop training, but literature baselines mixed into main tables require matched local controls. |
| 2610.07247 Learning What to Distill: Bilevel Top-K Token Selection for Self-Distillation in Large Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Soft-OR utility, softplus inner threshold, and soft masks form approximate bilevel selection; detach threshold and use full-vocabulary forward KL. Final checkpoint at 200 steps. TIP comparison also changes divergence and does not isolate selection. |
| 2610.07212 Reward-Driven Learning under Prompt-Level Differential Privacy | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DP-GRPO clips each prompt group as one DP-SGD record with one-update GRPO record loss. Privacy accounting and actual Qwen 1.5B LoRA go beyond a noise demonstration. |
| 2610.07177 CLM-as-a-Judge: Evaluating an Open Contrastive Decision Model on Public Judge Benchmarks | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Public judge evaluation and calibration cascade separate calibration/eval, use both orders and constant controls, and expose cost–risk. arXiv lists ancillary pins, splits, predictions, and judges/run_pass/metrics/analysis source. Hosted strong judge revision cannot be pinned; no execution, and one attachment failed in browser. |
| 2610.07118 AMBER: Training Long-Horizon Web Agents through Append-Only Memory | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Append-only two-turn memory/action share terminal GRPO advantage with template-token masking. Experiments initialize with SFT, not fully SFT-free, and require a web environment. |
| 2610.07043 TRIAGE: Direction-Aware Mismatch Stabilization of Native NVFP4 Reinforcement Learning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TRIAGE gates negative-advantage, negative-gap tokens by segment while normalizing response update mass, and applies bounded pseudo-Huber repair to positive advantages. Actual NVFP4 learner/sampler experiments; A100 emulation cannot claim native throughput. |
| 2610.06993 DART-ES: Difficulty-Aware Reweighting and Targeted Replay for Fine-Tuning LLMs with Evolution Strategies | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DART-ES estimates difficulty with an EMA of perturbation-population accuracy, and uses mastery thresholds and zero-success patience to replay rare solvable items. Actual LLM math ES comparisons; historical accuracy oracle cannot enter evaluation agents. |
| 2610.06947 FactorBench: A Portfolio-Aware Benchmark for Automated Factor Mining | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Public evaluation: symbolic/Python factor to date×asset signal contract in five markets, train 2017–21, validate 2022–23, test 2024–25, with portfolio and cost backtests. Official data/evaluator/mining can support specialized automated research, not universal LLM ability or live A/B. |
| 2610.06892 Axiom Satisfiability of Linear Rewards in Alignment | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LP3 jointly minimizes candidate slack and linear reward-margin violations on finite preference graphs via LP rather than ILP. Real Habermas preferences, no policy training or guarantee of axioms on unseen responses. |
| 2610.06830 MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Group-normalize quality, cost, and latency separately before combining, and use prefix-probed marginal answer quality for stage credit. Training latency is a proxy, not wall-clock reward. |
| 2610.06829 CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Nonnegative conformal self-verification bonus and URL-stratified return advantages alter credit. Report WAI training, VWA bank transfer, and OM2W without policy training separately. |
| 2610.06804 Sharpen Without Search: On-Policy Distillation of Sequence-Level Power Distribution | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. The Rényi gradient is weighted MLE under a teacher power distribution times the student distribution, with SMC resampling from student proposals. GSM8K training does not read reference answers; retain the entire SMC pipeline. |
| 2610.06782 T-Search: An Open Agentic Retriever and Playground for Hard Multi-Step Search | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Public retrieval-agent interface: frozen-corpus search/save_and_advance/finalize tools and multi-turn evidence state for TRuST/SynthComp. Official Apache-2.0 harness readable. Final-set Recall@10 and generative-agent trajectory recall differ; budget all three fusion calls. |
| 2610.06729 Improving Diversity in LLM Short Story Generation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DivLM's CPT instruction residual recovery and group genre/tone/style/entity diversity reward define two-stage training, requiring real judges and corpora. |
| 2610.06671 Learning What to Imitate: Entropy-Aware Distribution Mixing | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ConMix/GeoMix use normalized student Top-K entropy to set teacher contribution for sampled SFT and forward KL. AMiD already supplied both mixtures; entropy scheduling is new. |
| 2610.06648 Representation-Space MMD for Diffusion Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Use RBF-MMD over frozen DLM token features, sample discrete outputs with one denoiser pass, then group REINFORCE and differentiate after continuous bootstrap. OWT/GSM8K and 16B DMax experiments. |
| 2610.06647 LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Accumulate randomized gradient sketches directly, apply low-rank updates and predicted-KL step sizes. Real multiscale PPO; H100 results cannot be extrapolated to A100. |
| 2610.06637 Long-Horizon Textual World Modeling through Structured Reasoning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Real encode-transition-decode rollout trees assign sibling-group credit, jointly training predictive gain and intermediate rewards without gold textual states. |
| 2610.06446 The Assistance Dilemma: Learning to Teach via Multi-Turn Reinforcement Learning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Tutor has no gold solution. Masked near-transfer post-test retains only student tokens, with correctness/no-handover gates, frozen students, and cross-benchmark evaluation. |
| 2610.06439 Better Call Reward: Reward Hacking as Strategic Abstention in Legal Reasoning Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Public reward-hacking diagnostics: 16 LegalBench tasks separate surface proxy, parseable correctness, CTS, and rolling GDM. Apache-2.0 repository has notebooks and results. Deliberately bad proxy training detects format/abstention collapse, not capability gains. |
| 2610.05935 ThunderSyncRL: Lossless Acceleration of Agentic Reinforcement Learning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ThunderSyncRL streams fixed-policy-version GRPO dual-gradient accumulation and delayed token-normalized OPD before one optimizer barrier, preserving mathematical updates. B200/H200 checks and real agent training; stale asynchronous gradients are not a substitute. |
| 2610.05481 CodeForge-MA: Execution-Verified Multi-Agent Learning with Language-Conditioned LoRA for Multilingual Code Generation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. PPO+SFT is standard; language-conditioned sparse Mixture-of-LoRA gating is the executable addition. Qwen 1.8B and 72B API tests. Preserve routing and auxiliary constraints instead of renaming PPO. |
| 2610.05425 Unmentioned Checklist Findings Change How Reinforcement Learning Appears to Improve Chest Radiograph Report Checking | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Public checklist/reader-format robustness: image writer cannot see queried sentence, checker cannot see images. Matched-label pairs and three checkers isolate omission/order. Official code/formats/analysis accessible; CheXpert needs registration. No clinical or broad vision claim. |
| 2610.05373 Towards Unbiased On-Policy Distillation for Block Diffusion Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Un-OPD filters block decisions with aligned visible context, blends teacher and detached student logits at high teacher support, and reuses rollouts. Actual BDLM experiments. Title's unbiased does not guarantee arbitrary rollout reuse. |
| 2610.05367 AIProver: Agentic Auto-Formalization of Mathematical Research via Certificate-Driven Evolving Harness | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AIProver alternates model training and harness evolution, with SAM dual-view contrastive/CE, certificate RLSF, and independent fitness accept/rollback. Real formal verification required, not gold-proof fixtures. |
| 2610.05308 RubricArmor: Adversarial Evolution Improves LLM-Based Rubric Generation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. RubricArmor independently verifies flawed full-score attacks and iteratively patches rubrics with score-reduction and independent-review gates. 2K WildChat and actual GRPO on three policies. Preserve executable reward construction and verifiers, not string rules. |
| 2610.05303 ASCENT: Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ASCENT updates LoRA online only on verified success. A fixed initial teacher sees filtered self-hindsight and supervises a student without it via forward KL. No external gold; preserve strict prequential evaluation. |
| 2610.05282 Red-TTT: Test-Time Training for Automated Jailbreaking Large Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Red-TTT performs per-target test-time LoRA with softmax(beta reward) minus uniform risk-preference advantages, increasing beta and using elite replay. Fixed 120-query budget. Algorithm review only; no attacks executed. |
| 2610.05139 Hidden in the Comments: A Context-Injection Attack Surface in Code LLMs | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Public security evaluation: dual Flask prompt conditions, generation harness, five-component detector, and comment-position controls. Zenodo API confirmed open/published Manuscript_Codes.zip and checksums. Scaffolds are not paired and task types confound; detector references conditions rather than vulnerability gold, so no pure causal or true vulnerability-accuracy claims. |
| 2610.04950 How Should Teachers Be Prepared? RL on Student-Induced States for On-Policy Distillation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Prep-OPD first trains the teacher with RL on prefixes from a fixed initial student, then freezes it for OPD. DeepMath and eight math benchmarks. No gold solutions injected; preserve state distributions in both stages. |
| 2610.04918 Residual Visual Credit Optimization: Conserved Evidence Routing for Multimodal Reinforcement Learning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. RVCO scores evidence by likelihood changes after image-patch interventions, routes with IQR/entropy, and uses residual correction to preserve response credit mass. ViRL39K trains four VLM families; requires real visual interventions, not fixed token heuristics. |
| 2610.04882 ResOPD: Tail Residualization for Sparse On-Policy Distillation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ResOPD uses exact top-k mass and a tail bucket baseline, with coarse per-step gradients plus rare tail residual for unbiased estimation under full-support ancestral sampling only. Main math uses top-p=.95/top-k20, outside that condition; separate mechanism and results. |
| 2610.04851 CURIO: Curiosity-Driven Test-Time Learning for Open-Ended Discovery | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. CURIO trains a hidden-state transition predictor and caches pre-update residuals. Normalize by epoch and add bonuses only to sampled non-top-k tokens. Not guaranteed semantic novelty; encoder or predictor can collapse. Measured scientific discovery tasks. |
| 2610.04845 Which Preferences to Train On? End-to-End Multi-Objective Alignment with an Adversarial Preference Distribution | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MAESTRO uses DirUEG for an adversarial Dirichlet preference distribution, separates adversary/policy batches, and scalarizes vector-reward GAE by preference. Multiobjective HH/Beaver and summary faithfulness tests. |
| 2610.04596 DiffGate: Difficulty-Gated Teacher Guidance for On-Policy Distillation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DiffGate bounds the teacher–student log ratio with 2tanh(delta/2) and modulates teacher guidance only on failed traces according to group success. It trains 0.6B/1.7B math and code students with a frozen 8B teacher. |
| 2610.04344 Hierarchical Credit Assignment for RLVR on Fused Gromov-Wasserstein Geometry | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. HarA computes separate FGW barycenters for correct and incorrect hidden-state position graphs, assigns different positive and negative credit powers by span contribution to those barycenters, and multiplies GRPO advantage. Real multi-model and multi-benchmark experiments do not justify replacing FGW with Euclidean token scores. |
| 2610.04328 Bidirectional Preference Synthesis: Learning Prompt-Conditioned Preferences from Boundary Failures | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. BPS constructs offline preference data rather than a new loss. From partial failures, generate student corrections and safe, non-leaky achieved prompts. After cycle consistency, reverse the same response pair across two prompts as crossed anchors and apply standard DPO. IFBench and function-call tests cover mechanisms and downstream outcomes; same-teacher rankings are not independent generalization. |
| 2610.04299 Questioning the Questions: Sustaining Self-Evolution in Reasoning Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. R-Quest uses validity initialization and pairwise novelty gating of samples to change questioner–solver co-evolution. Two real model-family experiments are reported. The paper selects checkpoints by highest test benchmark mean per domain; an independent validation split is needed rather than inheriting these numbers as unbiased formal outcomes. |
| 2610.04132 PB-GRPO: Learning Socially Adaptive LLM Agents from Persona-Driven Simulation with Preference-Batched GRPO | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. PB-GRPO preserves prompt-local means but pools variance within latent preference buckets for the denominator. Actual Qwen7B experiments use 128 training and 100 unseen personas; this preliminary single-setting evidence warrants caution. |
| 2610.03984 Teaching Agents to Code Reliably | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Weighted coverage Shapley assigns credit to test assertions on mutation and sibling patches, gates positive trajectory advantage, and injects credit into assertion tokens. Gold tests are scorer-only. All 500 examples were first diagnosed, then split into 230 training and 270 holdout; do not claim the holdout was wholly unseen. |
| 2610.03190 Not Until the Evidence Says So: Teaching LLM Investigators When to Close a Case | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Nautil supplies evidence indexing and request tools, closed/open tasks, within-source balanced accuracy, and key-evidence removal controls. Official MIT code, data, and models are available. RLVR checkpoint selection used 22 of 64 test items, so isolate validation before treating any headline as leakage-free generalization. |
| 2610.03063 HARPO: Hallucination-Aware Reinforcement Learning for Faithful and Creative Language Generation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. HARPO trains a hallucination-span reward model and activates writing rewards only when no hallucinations are predicted, with a cosine creativity-to-fidelity curriculum. This is hierarchical multiobjective reward, not new GRPO. The gate does not guarantee hallucination-free outputs. |
| 2610.02986 OLMo-Detect: A Multi-Stage, Confounder-Controlled Benchmark for Membership Inference on Large Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. OLMo-Detect audits membership across pre-, mid-, and post-training, matching members and nonmembers by quality, time, and vocabulary. It includes 15 unsupervised and three supervised MIAs, separated dev/test, and official data and scoring. This measures membership and contamination risk, not stronger reasoning. |
| 2610.02856 Adaptive Mutual Distillation for Balanced Multi-Task Post-Training of Large Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AMD mutually distills MFT and smoothed-SFT branches. Three short global probes starting from the same state test increasing, decreasing, or preserving weights. Discard probe parameters and select direction per task on validation. Six benchmarks and several backbones were run; include probe costs. |
| 2610.02835 All Work And No Play Makes Jack a Dull Boy: Understanding and Preventing Catastrophic Strategy Collapse in RLVR | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MeshLearning explores with fixed offline coach-strategy prefixes containing no intermediate calculations or answers. Uniform KL balances growth in correct-trace policy log probability versus pre-RLVR reference. No coach at inference; actual math, science, and code training was run. |
| 2610.02828 FSPO: Policy-Consistent Risk and Pareto-Feasible Control for Budgeted LLM RL Post-Training | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. FSPO controls training actions using a policy-consistent risk-to-go ensemble, decision-conditioned calibration, and Pareto continuation certificates. Distinguish the 24-case ledger mechanism and real Qwen3B GRPO trainer study. Admit the mechanism, not fixed-ledger efficiency as LLM capability improvement. |
| 2610.02781 OPD Before RL: Warm-Starting Rubric-Based RL with On-Policy Distillation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. RP-OPD warms students with rubric-privileged teacher top-256 forward KL and then switches to rubric-reward GRPO across three rubric domains and multiple students. HealthBench reports overlapping and non-overlapping tests separately; formal reproduction needs fully isolated data. |
| 2610.02700 Learning from Evolving Errors: Adaptive Iterative Repair for On-Policy Distillation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AIR-OPD starts at the first error anchor and retains repair blocks covering 1−epsilon correction probability mass. It distills privileged multiround teacher guidance with gamma-depth/rho-immediate-success normalized weights. References are training-branch only, and actual DAPO training was run. |
| 2610.02695 Test-time Calibration Learning for Large Language Model Reasoning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. TTCL builds pseudo-correctness and confidence targets from answer agreement, separates advantages for reasoning/answer and confidence tokens, and freezes an EMA target cache each epoch. Unlabeled adaptation and evaluation on the same benchmark are transductive TTT, not independent unseen testing. |
| 2610.02614 What Is Lost in Post-Training? Default Collapse and the Loss of In-Context Steerability Across Diverse Perspectives | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SDM filters prompts by relevance judge, buckets stance with a second judge, and applies PPO with a population-level penalty for deviation from target distributions across dimensions. Two preliminary Llama8B settings do not establish per-prompt guarantees. |
| 2610.02444 Counterexample Generation via Per-Theorem Symbolic Verifiers: When Imitation Hurts and Reinforcement Repairs | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SymCE releases counterexample tasks, per-theorem Python verifiers, structural schemas, and dense/sparse rewards for RLVR/Evolve. Author repository includes corpus/splits, oracle/sandbox, runner, and item results. Schemas derive from canonical witnesses; audit possible answer-shape leakage and isolate arbitrary Python execution. |
| 2610.02359 Lexicographic Multi-Objective On-Policy Distillation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. LMOPD routes to the expert for the highest-priority unsatisfied reward, then applies one ordered projection to centered log-policy corrections for a virtual teacher. Real multi-expert experiments use private training data; one nonorthogonal projection cannot guarantee all constraints or no reward regression. |
| 2610.02324 Slow-Fast Multi-Teacher On-Policy Distillation for Capability Preservation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SF-MOPD maintains an EMA slow student, uses slow–fast log-probability directions to remove teacher conflict components, and applies JSD distillation with original MOPD regularization. QwenVL2/4/8B ran multiple domains. Use real EMA parameters rather than fixed references. |
| 2610.02206 KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. KaliBench releases 8,504 query–command pairs, 1,642 tools, and runtime-free shlex scoring under three information conditions with alias-aware arguments. This is offline command-string scoring, not real execution. Do not execute commands automatically against live targets. |
| 2610.02191 The Missing Primitive: Diagnosing and Repairing Mathematical Reasoning in Large Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Absorb gives the teacher mathematical primitives only in training. A one-sided reverse-KL clamp limits suppression of student preferences while retaining positive guidance. Prim/HLE/HMMT/Omni were tested; gold primitives must not reach evaluated students. |
| 2610.02190 Trust the Direction, Search the Step: Zero-and-First-Order Methods for LLM Fine-Tuning | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ZFO uses the base first-order optimizer direction and two same-batch positive/negative forwards to estimate second and third derivatives. Taylor/Padé interval step selection includes a singularity fallback. Actual LLM tuning and overhead were measured; match extra forward budgets. |
| 2610.02140 Finetuning with Sampling: SFT Learns Better Than You Think | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Projection Sampling uses expert-conditioned proposals and blockwise Metropolis–Hastings to keep informative traces close to the base distribution before SFT. Actual chemistry, math, and open-domain training was run. Finite MCMC is not exact sampling from an ideal projection. |
| 2610.02019 Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ATPO adjusts Tversky alpha/beta from EMA log-ratio error feedback of FN and FP, fixing their sum at two. It controls global or per-class multilabel reward. Actual VideoVLM training supports a new adaptive reward controller, not a renamed GRPO. |
| 2610.01955 Do Your Own Research: Learning to Forecast by Learning to Search | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. prime-forecast releases tool environments, historical questions, and rollout logs, with point-in-time filtering, blocked market prices, Brier scoring, and abstention imputation. Split train/test by resolution date. Live search and API filters are not static, and perfect filtering on a small sample is not leakage proof. |
| 2610.01892 Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SSR is not ordinary tool use: batched teacher-forced mean-logprob softmax chooses among six shared reasoning templates, then inserts full text before generating an action. SFT uses selector plus action NLL. GRPO uses one categorical selector ratio plus action-token ratios, with no token training on inserted templates. Appendix B detaches rollout competition scores; only first-step probabilities are exact and later ones approximate. Qwen3VL2B/4B run seven multimodal search tasks. Include candidate prefill in efficiency and retain the original post-training abstract-screen miss corrected by full text. |
| 2610.01670 Do MLLM Judges Judge the Edit? Auditing Bias in Image Editing Evaluation with Verified Quality Preservation | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. EditJudgeBias independently audits judge bias through quality-preserving cues, sham/retest noise baselines, swapped order, and paired statistics. Official cue/judge/metric CLI can test new judges, but missing full images, human labels, and author I2EBench outputs prevent exact reproduction. |
| 2610.01560 AURAL: Adaptive Latent Reasoning with Joint Chunk for Speech Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. AURAL models latent chunks jointly with diagonal plus low-rank GMM covariance, CSA, and differentiable two-pass scheduled sampling, then uses latent-coordinate-normalized ratios for GRPO. Preserve the full latent model when testing measured speech/text benchmarks and latency. |
| 2610.01316 What Wins a Vote? Formatting, Length, and Lexical Diversity in the French Compar:IA LLM Arena | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Public preference analysis truncates dialogue at voting time and fits joint style/length/vocabulary Bradley–Terry with within-model controls. Official code and derived data are readable; original French ComparIA is gated. There are 137,293 decisive votes over 116 models and 127,092 complete cases. Adjusted rank is not proven closer to capability, and observational association does not establish style causality. |
| 2610.01278 SCOPE-AD: Sequential cost-aware ordinal-belief planning with energy-based models for diagnostic agents | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. SCOPE-AD uses ordinal belief from observations and an offline soft Bellman EBM planning teacher, then distills with masked KL to Qwen. ADNI patient splits are 1,552/333/333; actual unobserved values are hidden. This is medical offline planning, not general OPD or clinical deployment. |
| 2610.01238 Mixture-Trained Merging for Unified Multi-Objective Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. MTM trains mixture-biased soft experts, chooses branch blends with multiobjective Bayesian optimization on the merge simplex, and reuses merged checkpoints iteratively. Real Qwen4B/OLMo7B multi-domain and thinking-mode experiments require validation selection and accounting for search training. |
| 2610.01128 Grounding Large Language Models in DSGE Simulators for Policy Generation and Forecasting | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. The public DSGE language-agent environment offers deep cloned Snowdrop simulator state, action validation, rolling forecasts committing the first step only, and immutable history. Matched PPO/GRPO share initial conditions and shocks. The author repository is readable for domain simulations, not real policy advice or trading causality. |
| 2610.01076 GLoC-EHR: Evidence-Cited Clinical Reasoning over Global Context and Local EHR Events | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. GLoC-EHR adds global/local EHR memory interfaces, staged alignment, and evidence-only sufficiency loss. 128,350 patients use isolated 7:1:2 splits and external hospital evaluation. RL remains standard GRPO; admit the architecture and grounded training mechanism. Only teacher/scorer see gold. |
| 2610.00779 Effective Synthetic Data Curation Requires Group-Level Signals | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Group-level influence measures downstream training utility, using within-group gradient diversity to allocate expensive group-oracle calls and individual proxies for unaudited groups. Actual pretraining/RLVR ran; isolate references and test. |
| 2610.00767 Pre-training interventions, ex post facto: Grafting model beliefs across checkpoints | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. Learn a synthetic-document delta in a base checkpoint and add it to the post-trained model to reduce reality drift. Multiple models up to 284B were tested. This differs from prior GRAFT trajectory exchange. |
| 2610.00661 Exploring More, Reasoning Better: Stepwise Risk-Sensitive GRPO for Diffusion Language Models | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. StepRS-GRPO changes risk coefficients and transforms group advantages by denoising maskedness, preserving the original diffusion trainer. With binary reward this is state/problem reweighting only. Actual SDAR1.7/4B training has mass and RMS controls. |
| 2610.00574 Make Sparse Rewards Count: Density-Aware Reward Aggregation for Multi-Reward RL | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. DARA weights active groups per reward by capped sqrt(max-density/density), while Asym amplifies only positive advantages before batch normalization. Actual ToolRL/BFCL-v4 and math training was run. |
| 2610.00568 Emergent Unfaithfulness: How Alignment Training Causes Language Models to Silently Override Task Faithfulness | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. FaithConflict publishes 940 confirming/conflicting document pairs, FaithGap, and a judge with eight behavior and seven reasoning categories, across 22 checkpoints and prompt conditions. Official code/data are readable. Separate task faithfulness and factual correctness; parroting harmful fictional documents is not safety improvement. |
| 2610.00431 ChainLoRA: Geometry-Preserving Task Vector Merging for Continual Learning in LLMs | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. ChainLoRA initializes QR carrier chains with one-sided geometric regularization, then archives, finds common carriers by SVD, aligns by Procrustes, and merges shared/residual parts. Three real continual-learning datasets were tested. |
| 2610.00385 FAER: Auditable Utility-Aligned Trajectory Replay for Language Model Post-Training | Needs cost/scope reduction or delivery; not auto-queued | Not independently confirmed this round; not excluded | Accepted in review; core implementation and public verification pending. FAER-UTILITY fits gradient alignment and metadata replay weights on non-target calibration blocks, samples without replacement, and joins correctness only after freezing traces. Completed learner results use normalized alignment. Optimizer-aware virtual steps are extensions or diagnostics, not validated learner behavior. |
See reconciliation.json for per-entry ledger references, titles, and current status.