Free tools Windows power users keep installed
One-click scans. No signup required.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Not yet. Excess Separability is a carefully designed method for looking for benchmark contamination in a model’s internal residual-stream activations, including after supervised fine-tuning (SFT) and reinforcement learning (RL). However, the revised 2026 preprint by Florian Braun withdraws its initial nominally significant finding: after accounting for instability in its placebo baseline, none of four real-transformer audit arms supports a contamination verdict. The method is therefore best understood as a promising, variance-aware audit protocol—not proof that a model memorized a benchmark or that internal probes reliably survive RL post-training.
What problem is Excess Separability trying to solve?
Benchmark contamination occurs when evaluation questions, solutions, or close variants appear in a model’s training process. A contaminated model may score well because it has encountered the test material, rather than because it learned a capability that generalizes to unseen problems. That distinction affects scientific conclusions, leaderboard fairness, and decisions about whether a model is ready for deployment.
Traditional checks usually examine outputs, token probabilities, or comparisons with a separately trained reference model. Those signals can weaken after post-training. A model can preserve a high benchmark score while changing its output distribution enough to look less like a memorizer. Excess Separability asks a different question: do benchmark members and carefully matched non-members occupy unusually separable positions in the model’s internal residual stream?
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →The approach does not require access to the original training corpus or retraining a reference model. It does require access to intermediate activations, so it is aimed at open-weight or otherwise white-box models rather than ordinary output-only commercial APIs.
#1 Best Overall
- Read Before You Buy — No Video Output: These adapters support charging and USB 2.0 data transfer, but cannot transmit video signals. Except for standard USB webcams (which use USB data only), they are not compatible with HDMI/DisplayPort cables, video-capable USB-C hubs, or docking stations with video output.
- Convert USB-A Ports to USB-C: Designed to connect USB-C earphones, cables, flash drives, card readers, and other USB-C accessories to standard USB-A ports. Plug-and-play with no drivers or software required.
- Aluminum Alloy Housing: Built with a sturdy aluminum alloy shell that aids in heat dissipation and protects against daily wear and scratches. Designed to maintain a stable and secure connection.
- Compact & Travel-Friendly: The ultra-compact design allows the adapter to stay plugged into your device without blocking adjacent ports or adding bulk, reducing wear and tear on your original USB ports.
- 12-Month Warranty: Backed by a 12-month manufacturer warranty for peace of mind. Designed to meet strict quality control standards for reliable everyday performance.
How does the internal-activation method work?
At every transformer layer, information is carried through a residual stream. Excess Separability probes those vectors with a depth-profile score and compares the result with controls designed to remove ordinary differences between examples.
1. Build suspect and reference sets
The suspect set contains benchmark items whose possible training exposure is being tested. The reference set contains matched non-members. The protocol specifies a reference set twice as large as the suspect set, improving the stability of comparisons when the suspect sample is small.
2. Measure separability across depth
A probe is evaluated at multiple transformer layers rather than at one arbitrary depth. The resulting depth profile records how well the two groups can be distinguished as information moves through the network. A contrast in this profile is the candidate contamination signal.
3. Subtract a level-matched placebo
Surface properties such as length, formatting, domain, and wording can make two groups separable even when neither group was memorized. The protocol therefore constructs a level-matched placebo baseline: a comparison expected to have similar nuisance differences but no contamination label. The observed depth-profile score is recentred against that baseline.
4. Use label permutations for the null
Labels are repeatedly permuted to create a null distribution. This tests whether the observed contrast is larger than would be expected from arbitrary assignment of member and non-member labels. The permutation null is not the only source of uncertainty; the placebo estimate also varies across split seeds, which became decisive in the revised analysis.
Rank #2
- 5-in-1 USB-C Hub: Experience comprehensive connectivity featuring a Power Delivery input, two USB-A 2.0 ports, a USB-A 3.0 port, and an HDMI port. (Note: The USB-C power delivery input port is only for connecting an external wall charger to power your laptop and cannot power peripheral devices.)
- 90W Pass-Through Charging: Achieve optimal charging with 90W pass-through power to your laptop, supported by a total input of 100W, with the hub reserving 10W for operational efficiency. (Note: Wall charger not included.)
- Quick Data Transfers: Accelerate your productivity with rapid data transfers using a high-speed 5Gbps USB 3.0 port and two 480Mbps USB 2.0 ports.
- 4K HDMI Display: Enhance your visual experience with a hub capable of delivering 4K resolution at 30Hz in both mirror and extend modes. Please note that this hub is compatible with MacBook (macOS 12 and newer), Windows 10 and 11, ChromeOS, and laptops equipped with DP Alt Mode and Power Delivery. Note: This device is not compatible with Linux.
- What You Get: Anker USB-C Hub (5-in-1, 4K HDMI), welcome guide, 18-month warranty, and our friendly customer service.
5. Require a variance-aware decision
A small nominal p-value is not sufficient if the baseline used to centre the statistic is itself unstable. Excess Separability’s revision propagates uncertainty from the placebo baseline before deciding whether the evidence is stronger than chance.
Did the revised preprint find contamination in real transformers?
No. Across four contaminated-checkpoint audit arms, the placebo baseline was 1.30 to 1.56 times as variable as the permutation null in every arm. A result first reported with a nominal p=0.0075 became p=0.0745 after that additional variance was propagated. The revised paper states, “No arm shows contamination,” and issues no contamination verdict.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
This is a statistical correction, not a claim that contamination was absent from the checkpoints. The experiments may have involved contaminated models, but the protocol did not establish that its measured separability exceeded the uncertainty of its controls. A researcher should therefore report the result as inconclusive rather than as a positive detector finding or a clean bill of health.
Why does reinforcement learning make contamination harder to detect?
RL can change the relationship between memorized training examples and observable model behavior. In the ICLR 2026 study by Wang, Li, Ko and Zhang, brief Group Relative Policy Optimization (GRPO) markedly reduced the member/non-member separability used by many output- and likelihood-based detectors.
Stage I: contamination during the transition to a reasoning model
In Stage I, benchmark data entered during SFT as a base model was converted into a reasoning model. Clean GRPO was then applied. In that setup, clean SFT improved pass@1 by 11.30%. SFT contamination added 8.82% average pass@1 inflation across six benchmarks. After GRPO, the contaminated model still retained 7.14% average performance inflation, even though conventional detection signals became harder to distinguish.
Rank #3
- Sleek 7-in-1 USB-C Hub: Features an HDMI port, two USB-A 3.0 ports, and a USB-C data port, each providing 5Gbps transfer speeds. It also includes a USB-C PD input port for charging up to 100W and dual SD and TF card slots, all in a compact design.
- Flawless 4K@60Hz Video with HDMI: Delivers exceptional clarity and smoothness with its 4K@60Hz HDMI port, making it ideal for high-definition presentations and entertainment. (Note: Only the HDMI port supports video projection; the USB-C port is for data transfer only.)
- Double Up on Efficiency: The two USB-A 3.0 ports and a USB-C port support a fast 5Gbps data rate, significantly boosting your transfer speeds and improving productivity.
- Fast and Reliable 85W Charging: Offers high-capacity, speedy charging for laptops up to 85W, so you spend less time tethered to an outlet and more time being productive.
- What You Get: Anker USB-C Hub (7-in-1), welcome guide, 18-month warranty, and our friendly customer service.
Stage II: chain-of-thought contamination in an already capable model
Stage II applied chain-of-thought benchmark contamination to an advanced reasoning model. Most detectors approached random guessing after the subsequent training conditions. LiRA, a representative likelihood-ratio method, averaged 58.74% AUROC in this advanced-LRM CoT-contamination setting—only modestly above the 50% random-ranking baseline.
What mechanism did the study propose?
The authors link the collapse in separability to PPO-style importance sampling and clipping used in policy optimization. These mechanisms reweight and constrain updates, allowing the model’s performance to retain contamination-related gains while disrupting the output patterns that detectors rely on. Their conclusion is that RL can conceal SFT contamination evidence and that CoT contamination of advanced reasoning models can leave little detectable evidence for existing memorization-driven methods.
Does Excess Separability actually survive RL post-training?
The current evidence does not establish that it does. The method was designed to target a gap exposed by RL-fragility research because internal activations might retain information that outputs no longer reveal. Yet the revised real-transformer audits produced no statistically supported contamination verdict after the placebo variance was included.
That distinction matters. “The probe can be run after RL” is an access and engineering claim; “the probe reliably detects contamination after RL” is a validated-performance claim. The first is supported for models that expose intermediate activations. The second remains unproven by the revised preprint.
How does Excess Separability compare with other detector families?
No detector family should be treated as universally reliable after post-training. The practical differences are in where each method obtains its signal, what controls it needs, and whether its statistical decision has been replicated.
Rank #4
- Dual Converters, Infinite Potential：Includes 2× USB C male to USB A female adapters and 2× USB A male to USB C female adapters. Perfect for a wide range of uses—tablets with Bluetooth keyboards, expand USB ports on macbook, and more. Two different converters for all your daily needs
- Next-Level 10Gbps & 3A Charging: No more slow 480Mbps, this usb to usb c adapter has a transfer speed of up to 10Gbps, allowing you to do more transferring in less time. This usb adapter fits both USB A and USB C charger, supporting up to 3A fast charging
- Upgraded Exquisite Craftsmanship: With an aluminum alloy housing and metal connector, the usbc to usb adapter is extremely durable and sturdy. Rigorously tested to withstand more than 10,000 times of plugging and unplugging, ensuring long-lasting performance
- Broad Compatible: The usb c to usb adapter widely supports all USB C/ USB A devices like laptops, tablets, cellphones, car chargers, and phone chargers. Such as compatible with MacBook Pro/Air 2023/2022, Thunderbolt 4/3 Devices,Apple MagSafe Watch 9/8/7/SE/Ultra, iPad Pro 2022/2021, Samsung Galaxy S23/S20/S10, and iPhone 17/16/15 Pro. Plug and play
- Please Note: To reach 10Gbps speed, keep the cable under 3.3 ft. For USB A Male to USB C adapters, try flipping the USB C connector. USB C Male to USB A adapters support bidirectional 10Gbps transfer within 3.3 ft
| Method family | Signal location | Strength | Important limitation |
|---|---|---|---|
| Output-behavior tests | Generated text, answers, refusal or style patterns | Can run through an ordinary inference endpoint | RL and chain-of-thought post-training can erase member/non-member differences |
| Likelihood and reference-model methods | Token log-probabilities or ratios against a reference | Quantitative and familiar to benchmark auditors | Require probability access or a suitable reference; LiRA averaged 58.74% AUROC in the cited advanced-LRM CoT setting |
| Excess Separability | Residual-stream activations across layer depth | Controls matched nuisance variables and does not require the original corpus or reference-model retraining | Needs intermediate activations; revised audits found placebo variance 1.30–1.56 times the permutation-null variance and no contamination verdict |
| Private-holdout and release controls | Evaluation design rather than a post hoc detector | Prevents exposure and remains usable for closed models | Does not identify contamination that already occurred |
What access and data are required to run the protocol?
You need a model that can return residual-stream activations at selected layers, a suspect benchmark set, a matched non-member reference set, and enough compute to collect activations repeatedly across split seeds and permutations. An output-only API that returns text without hidden states cannot run the protocol as described.
- Model access: white-box weights or an inference interface exposing intermediate residual states.
- Evaluation data: benchmark members and matched non-members, with the reference set twice the suspect-set size.
- Matching metadata: controls for length, formatting, domain, and other surface differences that could create separability.
- Statistical budget: multiple split seeds, a level-matched placebo, and enough label permutations to estimate the null distribution.
- Reporting plan: preserve raw depth profiles, placebo variability, permutation results, and corrected p-values rather than reporting only the smallest nominal value.
How should an auditor implement an activation-probe audit?
- Freeze the model version. Record the checkpoint, tokenizer, layer indexing, decoding-independent activation settings, and every post-training stage. Intermediate checkpoints are especially valuable because RL can alter detector signals between releases.
- Define contamination hypotheses before probing. Specify which benchmark split is suspect, what counts as a near-duplicate, and which non-member items are eligible as controls. Do not select controls after inspecting activation results.
- Construct matched sets. Match length, formatting, subject area, difficulty, and other visible properties. Keep the reference set at twice the suspect-set size where feasible.
- Extract residual-stream activations. Collect activations at the same positions and layers for every item. Keep preprocessing identical across groups and preserve the complete depth profile.
- Compute the probe and placebo. Estimate separability for the suspect/reference labels, then estimate a level-matched placebo that captures nuisance-driven separability. Recentre the observed score against that placebo.
- Generate the permutation null. Randomly permute labels many times while retaining the same examples and split structure. Compare the corrected statistic with this null, not with an uncalibrated threshold.
- Repeat across split seeds and checkpoints. A single favorable split can produce a misleading p-value when the baseline is noisy. Report the distribution of placebo estimates and corrected statistics across seeds.
- Issue a qualified decision. Say “evidence supports contamination” only when the corrected result is stable across prespecified splits and controls. Otherwise report “inconclusive,” not “clean.”
What common mistakes can invalidate the result?
- Confusing separability with memorization: Topic, length, or formatting differences can separate groups without any benchmark exposure.
- Using only the permutation variance: The revised preprint shows why uncertainty in the placebo baseline must also be propagated.
- Cherry-picking a layer or split: Searching many depths and seeds but reporting the best one inflates false positives.
- Assuming RL preserves detector signals: GRPO and related updates can conceal output and likelihood evidence while retaining performance inflation.
- Generalizing from one checkpoint: A result on an SFT checkpoint does not establish behavior after GRPO, PPO, or later reasoning training.
- Calling a non-significant test proof of no contamination: The revised result means the audit did not resolve the question; it does not certify the training history.
What should benchmark maintainers do while detectors remain fragile?
Evaluation design is the safer near-term control. Keep private holdouts out of training pipelines, use rolling release cutoffs, and document exactly when benchmark items became public. For open development, publish or retain intermediate checkpoints so auditors can examine changes introduced by SFT and RL instead of seeing only a final model.
Contamination-robust protocols should include fresh items, structurally similar but unseen tasks, and tests that cannot be solved by recalling a public chain of thought. Closed-model providers should disclose enough evaluation and checkpoint information for independent auditing, because an activation probe cannot compensate for an endpoint that exposes only final text.
What is the defensible conclusion in 2026?
Excess Separability is an important proposal because it addresses a real weakness in output-based contamination detection: RL can preserve inflated benchmark performance while hiding conventional signals. Its residual-stream design, nuisance-controlled placebo, doubled reference set, and permutation null are useful ingredients for future audits.
But the revised primary evidence is deliberately cautious. Once placebo-baseline sampling variance was included, the only nominally significant result crossed the usual threshold, from p=0.0075 to p=0.0745, and all four real-transformer audit arms were reported without a contamination verdict. The method therefore survives as a research direction and a protocol template, not as a demonstrated post-RL detector.
Best Value
- 5-in-1 Connectivity: Equipped with a 4K HDMI port, a 5 Gbps USB-C data port, two 5 Gbps USB-A ports, and a USB C 100W PD-IN port. Note: The USB C 100W PD-IN port supports only charging and does not support data transfer devices such as headphones or speakers.
- Powerful Pass-Through Charging: Supports up to 85W pass-through charging so you can power up your laptop while you use the hub. Note: Pass-through charging requires a charger (not included). Note: To achieve full power for iPad, we recommend using a 45W wall charger.
- Transfer Files in Seconds: Move files to and from your laptop at speeds of up to 5 Gbps via the USB-C and USB-A data ports. Note: The USB C 5Gbps Data port does not support video output.
- HD Display: Connect to the HDMI port to stream or mirror content to an external monitor in resolutions of up to 4K@30Hz. Note: The USB-C ports do not support video output.
- What You Get: Anker 332 USB-C Hub (5-in-1), welcome guide, our worry-free 18-month warranty, and friendly customer service.
Frequently Asked Questions
What is benchmark contamination in machine learning?
It is exposure of evaluation questions, answers, or close variants during training or post-training. The resulting score can overstate generalization because the model may be recalling material rather than solving an unseen task.
Why can a p-value change after a detector is corrected?
A p-value depends on the full uncertainty model. If the baseline used to centre a statistic varies across data splits, ignoring that variation makes the null distribution too narrow and the apparent evidence too strong.
Can I test a closed AI service with Excess Separability?
Not as specified. The protocol needs intermediate residual-stream activations, while a typical closed endpoint exposes only prompts, outputs, and sometimes token probabilities.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11What evidence would strengthen the method’s claim?
Independent replications across models and checkpoints, prespecified split seeds, matched clean controls, and variance-aware results that remain significant after placebo uncertainty is propagated would provide stronger support.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
